From 5ce016bb1f2f205e192cc61cf47422044004d2a0 Mon Sep 17 00:00:00 2001 From: Mohamed Boudra Date: Sun, 18 Oct 2026 09:02:00 +0200 Subject: [PATCH] feat(voice): add a Voice settings page with provider choice and a local model catalog Each host gets a Voice page under host settings. Dictation and voice mode each pick a provider (Local or OpenAI) and a model. Provider settings sit in one group of collapsible rows: local speech-to-text and text-to-speech list every catalog model with its size, languages and description before download, and OpenAI takes a write-only API key and an optional base URL. The local catalog grows from Parakeet/Kokoro to Whisper, SenseVoice, Dolphin, GigaAM, Parakeet Japanese, multilingual Kokoro, Kitten and Piper voices. Each entry declares its sherpa engine family, so the recognizer and TTS build the matching model config and the downloader verifies the files that family needs. Kitten v0.8 is left out: it crashes the pinned sherpa-onnx 1.12.28. Voice mode is now opt-in. The default lived only in code (true), never in config.json, so hosts that never toggled it find it off after upgrading. OpenAI speech-to-text resolved one model for both features, with voice mode's choice taking precedence, so dictation silently used voice mode's model. The resolved config now carries a model per feature; STT_MODEL still overrides both. Rows revealed under a collapsible settings row are a nested density level set by the container: tighter padding, smaller text, and xs controls, with controls taking their size from the row instead of each caller. --- docs/design.md | 4 +- .../components/settings/collapsible-row.tsx | 46 +- .../settings/headings/settings-info-tip.tsx | 28 +- .../app/src/components/settings/index.tsx | 23 +- .../settings/row-density.browser.test.tsx | 77 ++ .../src/components/settings/row-density.tsx | 57 ++ packages/app/src/components/ui/button.tsx | 5 +- .../components/ui/control-geometry.test.ts | 9 + .../app/src/components/ui/control-geometry.ts | 23 +- .../app/src/components/ui/control-size.tsx | 23 + .../src/components/ui/dropdown-trigger.tsx | 5 +- packages/app/src/components/ui/form-field.tsx | 24 +- packages/app/src/i18n/resources/ar.ts | 70 ++ packages/app/src/i18n/resources/en.ts | 70 ++ packages/app/src/i18n/resources/es.ts | 72 ++ packages/app/src/i18n/resources/fr.ts | 71 ++ packages/app/src/i18n/resources/ja.ts | 70 ++ packages/app/src/i18n/resources/ko.ts | 70 ++ packages/app/src/i18n/resources/pt-BR.ts | 71 ++ packages/app/src/i18n/resources/ru.ts | 72 ++ packages/app/src/i18n/resources/zh-CN.ts | 70 ++ packages/app/src/screens/settings-screen.tsx | 5 + .../settings/voice/use-local-speech-models.ts | 71 ++ .../src/screens/settings/voice/voice-page.tsx | 703 ++++++++++++++++++ .../settings/voice/voice-settings.test.ts | 126 ++++ .../screens/settings/voice/voice-settings.ts | 173 +++++ packages/app/src/utils/host-routes.ts | 1 + packages/app/vitest.config.ts | 2 + packages/client/src/daemon-client.ts | 29 + packages/protocol/src/messages.ts | 21 + packages/protocol/src/speech/rpc-schemas.ts | 165 ++++ .../authorization/operation-permissions.ts | 6 + packages/server/src/server/bootstrap.ts | 27 +- .../src/server/daemon-client.e2e.test.ts | 4 +- .../src/server/daemon-config-store.test.ts | 108 +++ .../server/src/server/daemon-config-store.ts | 82 +- packages/server/src/server/session.ts | 22 +- .../speech/speech-models-session.test.ts | 106 +++ .../session/speech/speech-models-session.ts | 124 +++ .../server/speech/providers/local/config.ts | 6 +- .../speech/providers/local/model-library.ts | 137 ++++ .../speech/providers/local/models.test.ts | 33 +- .../server/speech/providers/local/models.ts | 18 +- .../providers/local/sherpa/model-catalog.ts | 488 +++++++++++- .../local/sherpa/model-downloader.ts | 28 + .../local/sherpa/sherpa-offline-recognizer.ts | 59 +- .../providers/local/sherpa/sherpa-tts.test.ts | 49 +- .../providers/local/sherpa/sherpa-tts.ts | 77 +- .../local/sherpa/speech-download.e2e.test.ts | 17 +- .../speech/providers/local/worker-process.ts | 13 +- .../speech/providers/openai/config.test.ts | 17 + .../server/speech/providers/openai/config.ts | 39 +- .../speech/providers/openai/runtime.test.ts | 92 ++- .../server/speech/providers/openai/runtime.ts | 39 +- .../speech/speech-config-resolver.test.ts | 11 +- .../server/speech/speech-config-resolver.ts | 12 +- .../src/server/speech/speech-runtime.test.ts | 117 ++- .../src/server/speech/speech-runtime.ts | 198 +++-- .../src/server/speech/speech-settings.ts | 211 ++++++ .../src/server/test-utils/paseo-daemon.ts | 13 +- .../server/src/server/websocket-server.ts | 3 + public-docs/voice.md | 15 +- 62 files changed, 4136 insertions(+), 291 deletions(-) create mode 100644 packages/app/src/components/settings/row-density.browser.test.tsx create mode 100644 packages/app/src/components/settings/row-density.tsx create mode 100644 packages/app/src/components/ui/control-size.tsx create mode 100644 packages/app/src/screens/settings/voice/use-local-speech-models.ts create mode 100644 packages/app/src/screens/settings/voice/voice-page.tsx create mode 100644 packages/app/src/screens/settings/voice/voice-settings.test.ts create mode 100644 packages/app/src/screens/settings/voice/voice-settings.ts create mode 100644 packages/protocol/src/speech/rpc-schemas.ts create mode 100644 packages/server/src/server/session/speech/speech-models-session.test.ts create mode 100644 packages/server/src/server/session/speech/speech-models-session.ts create mode 100644 packages/server/src/server/speech/providers/local/model-library.ts create mode 100644 packages/server/src/server/speech/speech-settings.ts diff --git a/docs/design.md b/docs/design.md index e21270c85d..947c69ead6 100644 --- a/docs/design.md +++ b/docs/design.md @@ -70,7 +70,7 @@ the sidebar. Apply it as `backgroundColor`; setting `opacity` on the control als Sizes: `xs` for ultra-tight inline triggers. `sm` for any button sitting in a row. `md` is the page default. `lg` is reserved for large standalone CTAs. -Sizes are a shared contract across control kinds, defined once in `control-geometry.ts`: `xs` = 28px tall with `fontSize.sm` labels, `sm` = 32px with `fontSize.base`, `md`/`lg` = 44px with `fontSize.base`. `` (`packages/app/src/components/ui/segmented-control.tsx`) takes the same `xs`/`sm`/`md` sizes — a segmented control next to a ` diff --git a/packages/app/src/components/settings/row-density.browser.test.tsx b/packages/app/src/components/settings/row-density.browser.test.tsx new file mode 100644 index 0000000000..d2742e3172 --- /dev/null +++ b/packages/app/src/components/settings/row-density.browser.test.tsx @@ -0,0 +1,77 @@ +import React, { act } from "react"; +import { createRoot, type Root } from "react-dom/client"; +import { afterEach, describe, expect, it } from "vitest"; +import { Button } from "@/components/ui/button"; +import { SettingsCard, SettingsCollapsibleRow, SettingsRow } from "@/components/settings"; + +let root: Root | null = null; +let container: HTMLDivElement | null = null; + +afterEach(() => { + act(() => root?.unmount()); + container?.remove(); + root = null; + container = null; +}); + +function mountRows() { + container = document.createElement("div"); + document.body.appendChild(container); + root = createRoot(container); + act(() => + root?.render( + + + + + + + + + + + , + ), + ); +} + +function element(testID: string): HTMLElement { + const found = document.querySelector(`[data-testid="${testID}"]`); + if (!(found instanceof HTMLElement)) throw new Error(`Nothing rendered as ${testID}`); + return found; +} + +function heightOf(testID: string): number { + return element(testID).getBoundingClientRect().height; +} + +function titleFontSize(rowTestID: string, label: string): number { + const title = [...element(rowTestID).querySelectorAll("div")].find( + (node) => node.textContent === label && node.children.length === 0, + ); + if (!title) throw new Error(`No title ${label}`); + return Number.parseFloat(getComputedStyle(title).fontSize); +} + +describe("settings row density", () => { + it("lines up nested rows with and without a button, tighter than card rows", () => { + mountRows(); + + expect(heightOf("nested-with-button")).toBe(heightOf("nested-without-button")); + expect(heightOf("nested-with-button")).toBeLessThan(heightOf("card-row")); + expect(heightOf("group-toggle")).toBeGreaterThan(heightOf("nested-without-button")); + }); + + it("sizes nested text and controls one step below card rows", () => { + mountRows(); + + expect(heightOf("nested-button")).toBeLessThan(heightOf("card-button")); + expect(titleFontSize("nested-with-button", "Nested with a button")).toBeLessThan( + titleFontSize("card-row", "Card row"), + ); + }); +}); diff --git a/packages/app/src/components/settings/row-density.tsx b/packages/app/src/components/settings/row-density.tsx new file mode 100644 index 0000000000..d4bcdce33e --- /dev/null +++ b/packages/app/src/components/settings/row-density.tsx @@ -0,0 +1,57 @@ +import { createContext, useContext, useMemo, type ReactNode } from "react"; +import type { StyleProp, TextStyle, ViewStyle } from "react-native"; +import { StyleSheet } from "react-native-unistyles"; +import { CONTROL_HEIGHTS } from "@/components/ui/control-geometry"; +import type { ContainedControlSize } from "@/components/ui/control-size"; +import { settingsStyles } from "@/styles/settings"; + +// Rows revealed under a collapsible row belong to it, so they sit one step smaller than the rows +// of the card: tighter padding, smaller text, smaller controls. The container decides; a row +// never knows on its own that it is nested. +const NestedRowsContext = createContext(false); + +export function NestedSettingsRows({ children }: { children: ReactNode }) { + return {children}; +} + +export interface SettingsRowLayout { + row: StyleProp; + title: StyleProp; + hint: StyleProp; + /** The size of the controls in the row's trailing slot. */ + controlSize: ContainedControlSize; +} + +/** How a row lays out where it is rendered: in a card, or nested under a collapsible row. */ +export function useSettingsRowLayout(): SettingsRowLayout { + const nested = useContext(NestedRowsContext); + return useMemo( + () => + nested + ? { + row: [settingsStyles.row, styles.nestedRow], + title: [settingsStyles.rowTitle, styles.nestedText], + hint: [settingsStyles.rowHint, styles.nestedText], + controlSize: "xs", + } + : { + row: settingsStyles.row, + title: settingsStyles.rowTitle, + hint: settingsStyles.rowHint, + controlSize: "sm", + }, + [nested], + ); +} + +const styles = StyleSheet.create((theme) => ({ + // Every nested row is as tall as one holding an `xs` control, so rows with and without a + // trailing button line up. + nestedRow: { + paddingVertical: theme.spacing[2], + minHeight: CONTROL_HEIGHTS.tight + theme.spacing[2] * 2, + }, + nestedText: { + fontSize: theme.fontSize.sm, + }, +})); diff --git a/packages/app/src/components/ui/button.tsx b/packages/app/src/components/ui/button.tsx index 707ce60648..abcaf240d6 100644 --- a/packages/app/src/components/ui/button.tsx +++ b/packages/app/src/components/ui/button.tsx @@ -1,4 +1,5 @@ import { LoadingSpinner } from "@/components/ui/loading-spinner"; +import { useContainedControlSize } from "@/components/ui/control-size"; import { default as React, useCallback, @@ -162,7 +163,7 @@ const styles = StyleSheet.create((theme) => { export function Button({ children, variant = "secondary", - size = "md", + size: sizeProp, leftIcon, trailing, style, @@ -185,6 +186,8 @@ export function Button({ } >) { const [hovered, setHovered] = useState(false); + const containedSize = useContainedControlSize(); + const size = sizeProp ?? containedSize ?? "md"; const isDisabled = disabled || loading; let variantStyle: ViewStyle; diff --git a/packages/app/src/components/ui/control-geometry.test.ts b/packages/app/src/components/ui/control-geometry.test.ts index 0b0675596f..3544d9f940 100644 --- a/packages/app/src/components/ui/control-geometry.test.ts +++ b/packages/app/src/components/ui/control-geometry.test.ts @@ -81,6 +81,8 @@ describe("control geometry", () => { it("keeps field text sizing tied to control size", () => { const geometry = createControlGeometry(theme); + expect(geometry.fieldTextXs.fontSize).toBe(12); + expect(geometry.fieldTextXs.lineHeight).toBe(17); expect(geometry.fieldTextSm.fontSize).toBe(14); expect(geometry.fieldTextSm.lineHeight).toBe(20); expect(geometry.fieldTextMd.fontSize).toBe(14); @@ -95,6 +97,13 @@ describe("control geometry", () => { const geometry = createControlGeometry(theme); const borderWidth = theme.borderWidth[1]; + expect(geometry.fieldControlXs.minHeight).toBe(28); + expect( + geometry.fieldTextXs.lineHeight + + geometry.fieldControlXs.paddingVertical * 2 + + borderWidth * 2, + ).toBe(geometry.fieldControlXs.minHeight); + expect(geometry.fieldControlSm.minHeight).toBe(32); expect(geometry.fieldControlSm.paddingVertical).toBe(5); expect( diff --git a/packages/app/src/components/ui/control-geometry.ts b/packages/app/src/components/ui/control-geometry.ts index d56200887e..0a41f80039 100644 --- a/packages/app/src/components/ui/control-geometry.ts +++ b/packages/app/src/components/ui/control-geometry.ts @@ -2,7 +2,7 @@ import type { StyleProp, ViewStyle } from "react-native"; import { ICON_SIZE, type Theme } from "@/styles/theme"; export type ButtonControlSize = "xs" | "sm" | "md" | "lg"; -export type FieldControlSize = "sm" | "md"; +export type FieldControlSize = "xs" | "sm" | "md"; export type SegmentedControlSize = "xs" | "sm" | "md"; export type ControlInteractionPhase = "rest" | "hover" | "active"; @@ -188,8 +188,19 @@ export function resolveControlInteractionStyles( export function createControlGeometry(theme: Theme) { const controlBorderWidth = theme.borderWidth[1]; + const fieldTextXsLineHeight = fieldLineHeight(theme.fontSize.sm); const fieldTextSmLineHeight = fieldLineHeight(theme.fontSize.base); const fieldTextMdLineHeight = fieldLineHeight(theme.fontSize.base); + const fieldControlXs = { + minHeight: CONTROL_HEIGHTS.tight, + paddingHorizontal: theme.spacing[2], + paddingVertical: fieldVerticalPadding( + CONTROL_HEIGHTS.tight, + fieldTextXsLineHeight, + controlBorderWidth, + ), + borderRadius: theme.borderRadius.md, + }; const fieldControlSm = { minHeight: CONTROL_HEIGHTS.compact, paddingHorizontal: theme.spacing[3], @@ -210,6 +221,10 @@ export function createControlGeometry(theme: Theme) { ), borderRadius: theme.borderRadius.lg, }; + const fieldTextXs = { + fontSize: theme.fontSize.sm, + lineHeight: fieldTextXsLineHeight, + }; const fieldTextSm = { fontSize: theme.fontSize.base, lineHeight: fieldTextSmLineHeight, @@ -252,6 +267,10 @@ export function createControlGeometry(theme: Theme) { fontSize: theme.fontSize.sm, }, alert, + formTextInputXs: { + ...fieldControlXs, + ...fieldTextXs, + }, formTextInputSm: { ...fieldControlSm, ...fieldTextSm, @@ -264,8 +283,10 @@ export function createControlGeometry(theme: Theme) { ...fieldControlMd, ...fieldTextMd, }, + fieldControlXs, fieldControlSm, fieldControlMd, + fieldTextXs, fieldTextSm, fieldTextMd, controlRest: { diff --git a/packages/app/src/components/ui/control-size.tsx b/packages/app/src/components/ui/control-size.tsx new file mode 100644 index 0000000000..c6bcbbea78 --- /dev/null +++ b/packages/app/src/components/ui/control-size.tsx @@ -0,0 +1,23 @@ +import { createContext, useContext, type ReactNode } from "react"; + +/** The control sizes a container can ask of the controls placed in it. */ +export type ContainedControlSize = "xs" | "sm"; + +// A row decides how big its controls are, so a button or field placed in it needs no size of +// its own. A size passed to the control still wins. +const ControlSizeContext = createContext(null); + +export function ControlSizeProvider({ + size, + children, +}: { + size: ContainedControlSize; + children: ReactNode; +}) { + return {children}; +} + +/** The size the surrounding container asks for, or null outside one. */ +export function useContainedControlSize(): ContainedControlSize | null { + return useContext(ControlSizeContext); +} diff --git a/packages/app/src/components/ui/dropdown-trigger.tsx b/packages/app/src/components/ui/dropdown-trigger.tsx index 2fc1eb26ac..a276e40652 100644 --- a/packages/app/src/components/ui/dropdown-trigger.tsx +++ b/packages/app/src/components/ui/dropdown-trigger.tsx @@ -9,6 +9,7 @@ import { } from "@/components/ui/control-geometry"; import type { Theme } from "@/styles/theme"; import { DropdownMenuTrigger, type DropdownMenuTriggerProps } from "@/components/ui/dropdown-menu"; +import { useContainedControlSize } from "@/components/ui/control-size"; const ThemedChevronDown = withUnistyles(ChevronDown); @@ -32,10 +33,12 @@ interface DropdownTriggerProps extends Omit; +const FIELD_TEXT = { xs: "inputXs", sm: "inputSm", md: "inputMd" } as const satisfies Record< + FieldControlSize, + string +>; interface FieldProps { label: string; @@ -159,11 +169,13 @@ function assignTextInputRef( } export const FormTextInput = forwardRef( - function FormTextInput({ size = "md", style, onFocus, onBlur, editable, ...props }, ref) { + function FormTextInput({ size: sizeProp, style, onFocus, onBlur, editable, ...props }, ref) { const [focused, setFocused] = useState(false); + const containedSize = useContainedControlSize(); + const size = sizeProp ?? containedSize ?? "md"; const isDisabled = editable === false; - const chromeSizeStyle = size === "sm" ? formInputStyles.chromeSm : formInputStyles.chromeMd; - const inputSizeStyle = size === "sm" ? formInputStyles.inputSm : formInputStyles.inputMd; + const chromeSizeStyle = formInputStyles[FIELD_CHROME[size]]; + const inputSizeStyle = formInputStyles[FIELD_TEXT[size]]; const splitStyle = useMemo(() => splitFormTextInputStyle(style), [style]); const setInputRef = useCallback( (node: EditingTextInputHandle | null) => { @@ -262,6 +274,9 @@ const formInputStyles = StyleSheet.create((theme) => { chrome: { backgroundColor: theme.colors.surface2, }, + chromeXs: { + ...geometry.fieldControlXs, + }, chromeSm: { ...geometry.fieldControlSm, }, @@ -289,6 +304,9 @@ const formInputStyles = StyleSheet.create((theme) => { outlineColor: "transparent", outlineWidth: 0, }, + inputXs: { + ...geometry.fieldTextXs, + }, inputSm: { ...geometry.fieldTextSm, }, diff --git a/packages/app/src/i18n/resources/ar.ts b/packages/app/src/i18n/resources/ar.ts index 4adf945b61..627096b2a0 100644 --- a/packages/app/src/i18n/resources/ar.ts +++ b/packages/app/src/i18n/resources/ar.ts @@ -2106,6 +2106,7 @@ export const ar: TranslationResources = { connections: "اتصالات", agents: "Agents", metadata: "البيانات الوصفية", + voice: "الصوت", workspaces: "Workspaces", providers: "مقدمي الخدمات", usage: "الاستخدام", @@ -2128,6 +2129,75 @@ export const ar: TranslationResources = { docs: "الوثائق", saveError: "تعذر تحديث إنشاء البيانات الوصفية", }, + voice: { + offline: "اتصل بهذا المضيف لضبط الصوت.", + updateHost: "حدّث المضيف لضبط الصوت.", + saveError: "تعذّر تحديث إعدادات الصوت", + provider: "المزوّد", + providers: { + local: "محلي", + openai: "OpenAI", + }, + sttModel: "نموذج تحويل الكلام إلى نص", + ttsModel: "نموذج الصوت", + voice: "الصوت", + notDownloadedOption: "{{label}} (غير منزّل)", + save: "حفظ", + retry: "إعادة المحاولة", + dictation: { + title: "الإملاء", + enabled: "الإملاء", + enabledHint: "تحدّث في مربع الرسالة بدلًا من الكتابة", + }, + voiceMode: { + title: "وضع الصوت", + info: "يحوّل وضع الصوت كلامك إلى نص ويقرأ ردود الوكيل بصوت عالٍ.", + enabled: "وضع الصوت", + enabledHint: "تحدّث مع وكلائك دون استخدام يديك", + sttProvider: "تحويل الكلام إلى نص", + ttsProvider: "تحويل النص إلى كلام", + }, + openai: { + apiKey: "مفتاح API", + apiKeyHint: "يُحفظ على المضيف ولا يُعرض مرة أخرى.", + keyPlaceholder: "sk-...", + replaceKeyPlaceholder: "مفتاح جديد", + removeKey: "إزالة", + removeKeyTitle: "إزالة مفتاح OpenAI؟", + removeKeyMessage: "ستتوقف الميزات المضبوطة على OpenAI عن العمل حتى تضيف مفتاحًا.", + baseUrl: "عنوان URL الأساسي", + baseUrlHint: "للخوادم المتوافقة مع OpenAI. اتركه فارغًا لاستخدام OpenAI.", + missingKeyWarning: "تم اختيار OpenAI لكن لم يُضبط مفتاح API.", + }, + models: { + loading: "جارٍ تحميل النماذج...", + loadError: "تعذّر تحميل النماذج المحلية", + downloaded: "منزّل", + inUse: "قيد الاستخدام", + recommended: "موصى به", + download: "تنزيل", + downloading: "جارٍ التنزيل...", + downloadingPercent: "جارٍ التنزيل {{percent}}%", + extracting: "جارٍ الاستخراج...", + delete: "حذف", + deleteAction: "حذف {{label}}", + deleteTitle: "حذف {{label}}؟", + deleteMessage: "سيؤدي هذا إلى تحرير {{size}}. يمكنك تنزيله مرة أخرى لاحقًا.", + languages: "لغات {{label}}", + moreLanguages: "{{languages}} و{{count}} أخرى", + }, + providerSettings: { + title: "المزوّدون", + info: "النماذج وبيانات الاعتماد لكل مزوّد. تعمل النماذج المحلية على هذا المضيف وتعرض حجمها ولغاتها قبل التنزيل.", + localStt: "تحويل الكلام إلى نص محليًا", + localTts: "تحويل النص إلى كلام محليًا", + openai: "OpenAI", + downloaded_one: "تم تنزيل {{count}} نموذج", + downloaded_other: "تم تنزيل {{count}} نماذج", + configured: "تم الإعداد", + notConfigured: "غير مُعد", + }, + }, general: { title: "عام", sending: "الإرسال", diff --git a/packages/app/src/i18n/resources/en.ts b/packages/app/src/i18n/resources/en.ts index 9bb0460036..ac97e1c2f5 100644 --- a/packages/app/src/i18n/resources/en.ts +++ b/packages/app/src/i18n/resources/en.ts @@ -2169,6 +2169,7 @@ export const en = { connections: "Connections", agents: "Agents", metadata: "Metadata", + voice: "Voice", workspaces: "Workspaces", providers: "Providers", usage: "Usage", @@ -2253,6 +2254,75 @@ export const en = { docs: "Docs", saveError: "Unable to update metadata generation", }, + voice: { + offline: "Connect to this host to configure voice.", + updateHost: "Update the host to configure voice.", + saveError: "Unable to update voice settings", + provider: "Provider", + providers: { + local: "Local", + openai: "OpenAI", + }, + sttModel: "Speech-to-text model", + ttsModel: "Voice model", + voice: "Voice", + notDownloadedOption: "{{label}} (not downloaded)", + save: "Save", + retry: "Retry", + dictation: { + title: "Dictation", + enabled: "Dictation", + enabledHint: "Speak into the message box instead of typing", + }, + voiceMode: { + title: "Voice mode", + info: "Voice mode transcribes what you say and reads agent replies aloud.", + enabled: "Voice mode", + enabledHint: "Talk with your agents hands-free", + sttProvider: "Speech-to-text", + ttsProvider: "Text-to-speech", + }, + openai: { + apiKey: "API key", + apiKeyHint: "Saved on the host and never shown again.", + keyPlaceholder: "sk-...", + replaceKeyPlaceholder: "New key", + removeKey: "Remove", + removeKeyTitle: "Remove the OpenAI API key?", + removeKeyMessage: "Features set to OpenAI stop working until you add a key.", + baseUrl: "Base URL", + baseUrlHint: "For OpenAI-compatible servers. Leave empty to use OpenAI.", + missingKeyWarning: "OpenAI is selected but no API key is set.", + }, + models: { + loading: "Loading models...", + loadError: "Couldn't load local models", + downloaded: "Downloaded", + inUse: "In use", + recommended: "Recommended", + download: "Download", + downloading: "Downloading...", + downloadingPercent: "Downloading {{percent}}%", + extracting: "Extracting...", + delete: "Delete", + deleteAction: "Delete {{label}}", + deleteTitle: "Delete {{label}}?", + deleteMessage: "This frees {{size}}. You can download it again later.", + languages: "{{label}} languages", + moreLanguages: "{{languages}} and {{count}} more", + }, + providerSettings: { + title: "Providers", + info: "Models and credentials for each provider. Local models run on this host and list their size and languages before you download.", + localStt: "Local speech-to-text", + localTts: "Local text-to-speech", + openai: "OpenAI", + downloaded_one: "{{count}} model downloaded", + downloaded_other: "{{count}} models downloaded", + configured: "Configured", + notConfigured: "Not configured", + }, + }, general: { title: "General", sending: "Sending", diff --git a/packages/app/src/i18n/resources/es.ts b/packages/app/src/i18n/resources/es.ts index 9ed1e85572..e3e824b296 100644 --- a/packages/app/src/i18n/resources/es.ts +++ b/packages/app/src/i18n/resources/es.ts @@ -2154,6 +2154,7 @@ export const es: TranslationResources = { connections: "Conexiones", agents: "Agents", metadata: "Metadatos", + voice: "Voz", workspaces: "Workspaces", providers: "Proveedores", usage: "Uso", @@ -2176,6 +2177,77 @@ export const es: TranslationResources = { docs: "Documentación", saveError: "No se pudo actualizar la generación de metadatos", }, + voice: { + offline: "Conéctate a este host para configurar la voz.", + updateHost: "Actualiza el host para configurar la voz.", + saveError: "No se pudo actualizar la configuración de voz", + provider: "Proveedor", + providers: { + local: "Local", + openai: "OpenAI", + }, + sttModel: "Modelo de voz a texto", + ttsModel: "Modelo de voz", + voice: "Voz", + notDownloadedOption: "{{label}} (sin descargar)", + save: "Guardar", + retry: "Reintentar", + dictation: { + title: "Dictado", + enabled: "Dictado", + enabledHint: "Habla en el cuadro de mensaje en lugar de escribir", + }, + voiceMode: { + title: "Modo de voz", + info: "El modo de voz transcribe lo que dices y lee en voz alta las respuestas del agente.", + enabled: "Modo de voz", + enabledHint: "Habla con tus agentes sin usar las manos", + sttProvider: "Voz a texto", + ttsProvider: "Texto a voz", + }, + openai: { + apiKey: "Clave de API", + apiKeyHint: "Se guarda en el host y no se vuelve a mostrar.", + keyPlaceholder: "sk-...", + replaceKeyPlaceholder: "Nueva clave", + removeKey: "Quitar", + removeKeyTitle: "¿Quitar la clave de API de OpenAI?", + removeKeyMessage: + "Las funciones configuradas con OpenAI dejarán de funcionar hasta que añadas una clave.", + baseUrl: "URL base", + baseUrlHint: "Para servidores compatibles con OpenAI. Déjala vacía para usar OpenAI.", + missingKeyWarning: + "OpenAI está seleccionado, pero no hay ninguna clave de API configurada.", + }, + models: { + loading: "Cargando modelos...", + loadError: "No se pudieron cargar los modelos locales", + downloaded: "Descargado", + inUse: "En uso", + recommended: "Recomendado", + download: "Descargar", + downloading: "Descargando...", + downloadingPercent: "Descargando {{percent}} %", + extracting: "Extrayendo...", + delete: "Eliminar", + deleteAction: "Eliminar {{label}}", + deleteTitle: "¿Eliminar {{label}}?", + deleteMessage: "Esto libera {{size}}. Puedes volver a descargarlo más tarde.", + languages: "Idiomas de {{label}}", + moreLanguages: "{{languages}} y {{count}} más", + }, + providerSettings: { + title: "Proveedores", + info: "Modelos y credenciales de cada proveedor. Los modelos locales se ejecutan en este host e indican su tamaño e idiomas antes de descargarlos.", + localStt: "Voz a texto local", + localTts: "Texto a voz local", + openai: "OpenAI", + downloaded_one: "{{count}} modelo descargado", + downloaded_other: "{{count}} modelos descargados", + configured: "Configurado", + notConfigured: "Sin configurar", + }, + }, general: { title: "General", sending: "Envío", diff --git a/packages/app/src/i18n/resources/fr.ts b/packages/app/src/i18n/resources/fr.ts index 6563b1100e..5f1465ddbb 100644 --- a/packages/app/src/i18n/resources/fr.ts +++ b/packages/app/src/i18n/resources/fr.ts @@ -2153,6 +2153,7 @@ export const fr: TranslationResources = { connections: "Connexions", agents: "Agents", metadata: "Métadonnées", + voice: "Voix", workspaces: "Espaces de travail", providers: "Fournisseurs", usage: "Utilisation", @@ -2175,6 +2176,76 @@ export const fr: TranslationResources = { docs: "Documentation", saveError: "Impossible de mettre à jour la génération de métadonnées", }, + voice: { + offline: "Connectez-vous à cet hôte pour configurer la voix.", + updateHost: "Mettez à jour l'hôte pour configurer la voix.", + saveError: "Impossible de mettre à jour les réglages vocaux", + provider: "Fournisseur", + providers: { + local: "Local", + openai: "OpenAI", + }, + sttModel: "Modèle de transcription", + ttsModel: "Modèle de voix", + voice: "Voix", + notDownloadedOption: "{{label}} (non téléchargé)", + save: "Enregistrer", + retry: "Réessayer", + dictation: { + title: "Dictée", + enabled: "Dictée", + enabledHint: "Parlez dans la zone de message au lieu de taper", + }, + voiceMode: { + title: "Mode vocal", + info: "Le mode vocal transcrit ce que vous dites et lit à voix haute les réponses de l'agent.", + enabled: "Mode vocal", + enabledHint: "Parlez à vos agents sans les mains", + sttProvider: "Transcription", + ttsProvider: "Synthèse vocale", + }, + openai: { + apiKey: "Clé d'API", + apiKeyHint: "Enregistrée sur l'hôte, jamais réaffichée.", + keyPlaceholder: "sk-...", + replaceKeyPlaceholder: "Nouvelle clé", + removeKey: "Supprimer", + removeKeyTitle: "Supprimer la clé d'API OpenAI ?", + removeKeyMessage: + "Les fonctions réglées sur OpenAI cesseront de fonctionner jusqu'à l'ajout d'une clé.", + baseUrl: "URL de base", + baseUrlHint: "Pour les serveurs compatibles OpenAI. Laissez vide pour utiliser OpenAI.", + missingKeyWarning: "OpenAI est sélectionné, mais aucune clé d'API n'est définie.", + }, + models: { + loading: "Chargement des modèles...", + loadError: "Impossible de charger les modèles locaux", + downloaded: "Téléchargé", + inUse: "Utilisé", + recommended: "Recommandé", + download: "Télécharger", + downloading: "Téléchargement...", + downloadingPercent: "Téléchargement {{percent}} %", + extracting: "Extraction...", + delete: "Supprimer", + deleteAction: "Supprimer {{label}}", + deleteTitle: "Supprimer {{label}} ?", + deleteMessage: "Cela libère {{size}}. Vous pourrez le retélécharger plus tard.", + languages: "Langues de {{label}}", + moreLanguages: "{{languages}} et {{count}} autres", + }, + providerSettings: { + title: "Fournisseurs", + info: "Modèles et identifiants de chaque fournisseur. Les modèles locaux s'exécutent sur cet hôte et indiquent leur taille et leurs langues avant le téléchargement.", + localStt: "Reconnaissance vocale locale", + localTts: "Synthèse vocale locale", + openai: "OpenAI", + downloaded_one: "{{count}} modèle téléchargé", + downloaded_other: "{{count}} modèles téléchargés", + configured: "Configuré", + notConfigured: "Non configuré", + }, + }, general: { title: "Général", sending: "Envoi", diff --git a/packages/app/src/i18n/resources/ja.ts b/packages/app/src/i18n/resources/ja.ts index eb22dd663a..53f77465ac 100644 --- a/packages/app/src/i18n/resources/ja.ts +++ b/packages/app/src/i18n/resources/ja.ts @@ -2127,6 +2127,7 @@ export const ja: TranslationResources = { connections: "接続", agents: "エージェント", metadata: "メタデータ", + voice: "音声", workspaces: "ワークスペース", providers: "プロバイダー", usage: "使用状況", @@ -2149,6 +2150,75 @@ export const ja: TranslationResources = { docs: "ドキュメント", saveError: "メタデータ生成を更新できません", }, + voice: { + offline: "音声を設定するには、このホストに接続してください。", + updateHost: "音声を設定するには、ホストを更新してください。", + saveError: "音声設定を更新できませんでした", + provider: "プロバイダー", + providers: { + local: "ローカル", + openai: "OpenAI", + }, + sttModel: "音声認識モデル", + ttsModel: "音声モデル", + voice: "ボイス", + notDownloadedOption: "{{label}}（未ダウンロード）", + save: "保存", + retry: "再試行", + dictation: { + title: "音声入力", + enabled: "音声入力", + enabledHint: "入力する代わりにメッセージ欄に話しかけます", + }, + voiceMode: { + title: "ボイスモード", + info: "ボイスモードは話した内容を文字起こしし、エージェントの返答を読み上げます。", + enabled: "ボイスモード", + enabledHint: "ハンズフリーでエージェントと会話します", + sttProvider: "音声認識", + ttsProvider: "音声合成", + }, + openai: { + apiKey: "API キー", + apiKeyHint: "ホストに保存され、再表示されません。", + keyPlaceholder: "sk-...", + replaceKeyPlaceholder: "新しいキー", + removeKey: "削除", + removeKeyTitle: "OpenAI API キーを削除しますか？", + removeKeyMessage: "キーを追加するまで、OpenAI に設定された機能は動作しません。", + baseUrl: "ベース URL", + baseUrlHint: "OpenAI 互換サーバー用です。OpenAI を使う場合は空欄のままにします。", + missingKeyWarning: "OpenAI が選択されていますが、API キーが設定されていません。", + }, + models: { + loading: "モデルを読み込み中...", + loadError: "ローカルモデルを読み込めませんでした", + downloaded: "ダウンロード済み", + inUse: "使用中", + recommended: "おすすめ", + download: "ダウンロード", + downloading: "ダウンロード中...", + downloadingPercent: "ダウンロード中 {{percent}}%", + extracting: "展開中...", + delete: "削除", + deleteAction: "{{label}} を削除", + deleteTitle: "{{label}} を削除しますか？", + deleteMessage: "{{size}} が解放されます。後で再ダウンロードできます。", + languages: "{{label}} の対応言語", + moreLanguages: "{{languages}} ほか {{count}} 言語", + }, + providerSettings: { + title: "プロバイダー", + info: "各プロバイダーのモデルと認証情報です。ローカルモデルはこのホストで動作し、ダウンロード前にサイズと対応言語を確認できます。", + localStt: "ローカル音声認識", + localTts: "ローカル音声合成", + openai: "OpenAI", + downloaded_one: "{{count}} 個のモデルをダウンロード済み", + downloaded_other: "{{count}} 個のモデルをダウンロード済み", + configured: "設定済み", + notConfigured: "未設定", + }, + }, general: { title: "一般", sending: "送信", diff --git a/packages/app/src/i18n/resources/ko.ts b/packages/app/src/i18n/resources/ko.ts index afdb8974e9..4841f5f6c0 100644 --- a/packages/app/src/i18n/resources/ko.ts +++ b/packages/app/src/i18n/resources/ko.ts @@ -2116,6 +2116,7 @@ export const ko: TranslationResources = { connections: "연결", agents: "에이전트", metadata: "메타데이터", + voice: "음성", workspaces: "워크스페이스", providers: "프로바이더", usage: "사용량", @@ -2138,6 +2139,75 @@ export const ko: TranslationResources = { docs: "문서", saveError: "메타데이터 생성을 업데이트할 수 없습니다", }, + voice: { + offline: "음성을 설정하려면 이 호스트에 연결하세요.", + updateHost: "음성을 설정하려면 호스트를 업데이트하세요.", + saveError: "음성 설정을 업데이트할 수 없습니다", + provider: "제공자", + providers: { + local: "로컬", + openai: "OpenAI", + }, + sttModel: "음성 인식 모델", + ttsModel: "음성 모델", + voice: "목소리", + notDownloadedOption: "{{label}} (다운로드 안 됨)", + save: "저장", + retry: "다시 시도", + dictation: { + title: "받아쓰기", + enabled: "받아쓰기", + enabledHint: "입력하는 대신 메시지 상자에 말하세요", + }, + voiceMode: { + title: "음성 모드", + info: "음성 모드는 말한 내용을 텍스트로 옮기고 에이전트의 답변을 소리 내어 읽어 줍니다.", + enabled: "음성 모드", + enabledHint: "손을 쓰지 않고 에이전트와 대화하세요", + sttProvider: "음성 인식", + ttsProvider: "음성 합성", + }, + openai: { + apiKey: "API 키", + apiKeyHint: "호스트에 저장되며 다시 표시되지 않습니다.", + keyPlaceholder: "sk-...", + replaceKeyPlaceholder: "새 키", + removeKey: "제거", + removeKeyTitle: "OpenAI API 키를 제거할까요?", + removeKeyMessage: "키를 추가할 때까지 OpenAI로 설정된 기능이 작동하지 않습니다.", + baseUrl: "기본 URL", + baseUrlHint: "OpenAI 호환 서버용입니다. OpenAI를 사용하려면 비워 두세요.", + missingKeyWarning: "OpenAI가 선택되었지만 API 키가 설정되지 않았습니다.", + }, + models: { + loading: "모델을 불러오는 중...", + loadError: "로컬 모델을 불러올 수 없습니다", + downloaded: "다운로드됨", + inUse: "사용 중", + recommended: "추천", + download: "다운로드", + downloading: "다운로드 중...", + downloadingPercent: "다운로드 중 {{percent}}%", + extracting: "압축 해제 중...", + delete: "삭제", + deleteAction: "{{label}} 삭제", + deleteTitle: "{{label}}을(를) 삭제할까요?", + deleteMessage: "{{size}}가 확보됩니다. 나중에 다시 다운로드할 수 있습니다.", + languages: "{{label}} 지원 언어", + moreLanguages: "{{languages}} 외 {{count}}개", + }, + providerSettings: { + title: "제공자", + info: "제공자별 모델과 자격 증명입니다. 로컬 모델은 이 호스트에서 실행되며 다운로드 전에 크기와 지원 언어를 보여줍니다.", + localStt: "로컬 음성 인식", + localTts: "로컬 음성 합성", + openai: "OpenAI", + downloaded_one: "모델 {{count}}개 다운로드됨", + downloaded_other: "모델 {{count}}개 다운로드됨", + configured: "설정됨", + notConfigured: "설정 안 됨", + }, + }, general: { title: "일반", sending: "전송", diff --git a/packages/app/src/i18n/resources/pt-BR.ts b/packages/app/src/i18n/resources/pt-BR.ts index f89fb97bf0..5e519d81ec 100644 --- a/packages/app/src/i18n/resources/pt-BR.ts +++ b/packages/app/src/i18n/resources/pt-BR.ts @@ -2139,6 +2139,7 @@ export const ptBR: TranslationResources = { connections: "Conexões", agents: "Agentes", metadata: "Metadados", + voice: "Voz", workspaces: "Workspaces", providers: "Provedores", usage: "Uso", @@ -2161,6 +2162,76 @@ export const ptBR: TranslationResources = { docs: "Documentação", saveError: "Não foi possível atualizar a geração de metadados", }, + voice: { + offline: "Conecte-se a este host para configurar a voz.", + updateHost: "Atualize o host para configurar a voz.", + saveError: "Não foi possível atualizar as configurações de voz", + provider: "Provedor", + providers: { + local: "Local", + openai: "OpenAI", + }, + sttModel: "Modelo de fala para texto", + ttsModel: "Modelo de voz", + voice: "Voz", + notDownloadedOption: "{{label}} (não baixado)", + save: "Salvar", + retry: "Tentar novamente", + dictation: { + title: "Ditado", + enabled: "Ditado", + enabledHint: "Fale na caixa de mensagem em vez de digitar", + }, + voiceMode: { + title: "Modo de voz", + info: "O modo de voz transcreve o que você diz e lê em voz alta as respostas do agente.", + enabled: "Modo de voz", + enabledHint: "Converse com seus agentes sem usar as mãos", + sttProvider: "Fala para texto", + ttsProvider: "Texto para fala", + }, + openai: { + apiKey: "Chave de API", + apiKeyHint: "Salva no host e nunca mostrada novamente.", + keyPlaceholder: "sk-...", + replaceKeyPlaceholder: "Nova chave", + removeKey: "Remover", + removeKeyTitle: "Remover a chave de API da OpenAI?", + removeKeyMessage: + "Os recursos configurados com a OpenAI param de funcionar até você adicionar uma chave.", + baseUrl: "URL base", + baseUrlHint: "Para servidores compatíveis com a OpenAI. Deixe vazio para usar a OpenAI.", + missingKeyWarning: "A OpenAI está selecionada, mas nenhuma chave de API foi definida.", + }, + models: { + loading: "Carregando modelos...", + loadError: "Não foi possível carregar os modelos locais", + downloaded: "Baixado", + inUse: "Em uso", + recommended: "Recomendado", + download: "Baixar", + downloading: "Baixando...", + downloadingPercent: "Baixando {{percent}}%", + extracting: "Extraindo...", + delete: "Excluir", + deleteAction: "Excluir {{label}}", + deleteTitle: "Excluir {{label}}?", + deleteMessage: "Isso libera {{size}}. Você pode baixá-lo de novo depois.", + languages: "Idiomas do {{label}}", + moreLanguages: "{{languages}} e mais {{count}}", + }, + providerSettings: { + title: "Provedores", + info: "Modelos e credenciais de cada provedor. Os modelos locais rodam neste host e mostram tamanho e idiomas antes do download.", + localStt: "Fala para texto local", + localTts: "Texto para fala local", + openai: "OpenAI", + downloaded_one: "{{count}} modelo baixado", + downloaded_other: "{{count}} modelos baixados", + configured: "Configurado", + notConfigured: "Não configurado", + }, + }, general: { title: "Geral", sending: "Envio", diff --git a/packages/app/src/i18n/resources/ru.ts b/packages/app/src/i18n/resources/ru.ts index a1a70822b4..81108c36e0 100644 --- a/packages/app/src/i18n/resources/ru.ts +++ b/packages/app/src/i18n/resources/ru.ts @@ -2139,6 +2139,7 @@ export const ru: TranslationResources = { connections: "Подключения", agents: "Агенты", metadata: "Метаданные", + voice: "Голос", workspaces: "Рабочие пространства", providers: "Провайдеры", usage: "Использование", @@ -2161,6 +2162,77 @@ export const ru: TranslationResources = { docs: "Документация", saveError: "Не удалось обновить настройки генерации метаданных", }, + voice: { + offline: "Подключитесь к этому хосту, чтобы настроить голос.", + updateHost: "Обновите хост, чтобы настроить голос.", + saveError: "Не удалось обновить голосовые настройки", + provider: "Провайдер", + providers: { + local: "Локально", + openai: "OpenAI", + }, + sttModel: "Модель распознавания речи", + ttsModel: "Модель голоса", + voice: "Голос", + notDownloadedOption: "{{label}} (не скачана)", + save: "Сохранить", + retry: "Повторить", + dictation: { + title: "Диктовка", + enabled: "Диктовка", + enabledHint: "Говорите в поле сообщения вместо набора текста", + }, + voiceMode: { + title: "Голосовой режим", + info: "Голосовой режим расшифровывает вашу речь и зачитывает ответы агента вслух.", + enabled: "Голосовой режим", + enabledHint: "Общайтесь с агентами без рук", + sttProvider: "Распознавание речи", + ttsProvider: "Синтез речи", + }, + openai: { + apiKey: "API-ключ", + apiKeyHint: "Хранится на хосте и больше не показывается.", + keyPlaceholder: "sk-...", + replaceKeyPlaceholder: "Новый ключ", + removeKey: "Удалить", + removeKeyTitle: "Удалить API-ключ OpenAI?", + removeKeyMessage: + "Функции, настроенные на OpenAI, перестанут работать, пока вы не добавите ключ.", + baseUrl: "Базовый URL", + baseUrlHint: + "Для серверов, совместимых с OpenAI. Оставьте пустым, чтобы использовать OpenAI.", + missingKeyWarning: "Выбран OpenAI, но API-ключ не задан.", + }, + models: { + loading: "Загрузка моделей...", + loadError: "Не удалось загрузить локальные модели", + downloaded: "Скачана", + inUse: "Используется", + recommended: "Рекомендуется", + download: "Скачать", + downloading: "Скачивание...", + downloadingPercent: "Скачивание {{percent}}%", + extracting: "Распаковка...", + delete: "Удалить", + deleteAction: "Удалить {{label}}", + deleteTitle: "Удалить {{label}}?", + deleteMessage: "Освободится {{size}}. Позже её можно скачать снова.", + languages: "Языки {{label}}", + moreLanguages: "{{languages}} и ещё {{count}}", + }, + providerSettings: { + title: "Провайдеры", + info: "Модели и учётные данные каждого провайдера. Локальные модели работают на этом хосте; размер и языки видны до загрузки.", + localStt: "Локальное распознавание речи", + localTts: "Локальный синтез речи", + openai: "OpenAI", + downloaded_one: "Загружено моделей: {{count}}", + downloaded_other: "Загружено моделей: {{count}}", + configured: "Настроено", + notConfigured: "Не настроено", + }, + }, general: { title: "Основные", sending: "Отправка", diff --git a/packages/app/src/i18n/resources/zh-CN.ts b/packages/app/src/i18n/resources/zh-CN.ts index ca221d7d3a..d0b31445b9 100644 --- a/packages/app/src/i18n/resources/zh-CN.ts +++ b/packages/app/src/i18n/resources/zh-CN.ts @@ -2084,6 +2084,7 @@ export const zhCN: TranslationResources = { connections: "连接", agents: "Agents", metadata: "元数据", + voice: "语音", workspaces: "Workspaces", providers: "Providers", usage: "使用情况", @@ -2105,6 +2106,75 @@ export const zhCN: TranslationResources = { docs: "文档", saveError: "无法更新元数据生成设置", }, + voice: { + offline: "连接到此主机以配置语音。", + updateHost: "请更新主机以配置语音。", + saveError: "无法更新语音设置", + provider: "提供方", + providers: { + local: "本地", + openai: "OpenAI", + }, + sttModel: "语音转文字模型", + ttsModel: "语音模型", + voice: "音色", + notDownloadedOption: "{{label}}（未下载）", + save: "保存", + retry: "重试", + dictation: { + title: "听写", + enabled: "听写", + enabledHint: "对着消息框说话，无需打字", + }, + voiceMode: { + title: "语音模式", + info: "语音模式会转写你说的话，并朗读智能体的回复。", + enabled: "语音模式", + enabledHint: "免手动与智能体对话", + sttProvider: "语音转文字", + ttsProvider: "文字转语音", + }, + openai: { + apiKey: "API 密钥", + apiKeyHint: "保存在主机上，不会再次显示。", + keyPlaceholder: "sk-...", + replaceKeyPlaceholder: "新密钥", + removeKey: "移除", + removeKeyTitle: "移除 OpenAI API 密钥？", + removeKeyMessage: "在添加密钥之前，设为 OpenAI 的功能将无法使用。", + baseUrl: "基础 URL", + baseUrlHint: "用于兼容 OpenAI 的服务器。留空则使用 OpenAI。", + missingKeyWarning: "已选择 OpenAI，但未设置 API 密钥。", + }, + models: { + loading: "正在加载模型...", + loadError: "无法加载本地模型", + downloaded: "已下载", + inUse: "使用中", + recommended: "推荐", + download: "下载", + downloading: "正在下载...", + downloadingPercent: "正在下载 {{percent}}%", + extracting: "正在解压...", + delete: "删除", + deleteAction: "删除 {{label}}", + deleteTitle: "删除 {{label}}？", + deleteMessage: "这将释放 {{size}}。之后可以重新下载。", + languages: "{{label}} 支持的语言", + moreLanguages: "{{languages}} 等另外 {{count}} 种", + }, + providerSettings: { + title: "提供方", + info: "各提供方的模型和凭据。本地模型在此主机上运行，下载前会显示大小和支持的语言。", + localStt: "本地语音转文字", + localTts: "本地文字转语音", + openai: "OpenAI", + downloaded_one: "已下载 {{count}} 个模型", + downloaded_other: "已下载 {{count}} 个模型", + configured: "已配置", + notConfigured: "未配置", + }, + }, general: { title: "通用", sending: "发送", diff --git a/packages/app/src/screens/settings-screen.tsx b/packages/app/src/screens/settings-screen.tsx index 31f5058940..55b67819fa 100644 --- a/packages/app/src/screens/settings-screen.tsx +++ b/packages/app/src/screens/settings-screen.tsx @@ -36,6 +36,7 @@ import { Code2, Smartphone, Sparkles, + Mic, Blocks, Globe, PanelLeft, @@ -117,6 +118,7 @@ import { PluginSettingsContent } from "@/plugins/settings"; import { useInstalledPlugins } from "@/plugins/registry"; import { HostPluginsPage } from "@/screens/settings/plugins-page"; import { MetadataGenerationPage } from "@/screens/settings/metadata-generation-page"; +import { HostVoicePage } from "@/screens/settings/voice/voice-page"; import ProjectsScreen from "@/screens/projects-screen"; import ProjectSettingsScreen from "@/screens/project-settings-screen"; import { SETTINGS_DESKTOP_SIDEBAR_WIDTH, useIsCompactFormFactor } from "@/constants/layout"; @@ -234,6 +236,7 @@ const HOST_SECTION_ITEMS: HostSectionItem[] = [ { id: "pair-device", labelKey: "openProject.tiles.pairDevice.title", icon: Smartphone }, { id: "agents", labelKey: "settings.hostSections.agents", icon: Bot }, { id: "metadata", labelKey: "settings.hostSections.metadata", icon: Sparkles }, + { id: "voice", labelKey: "settings.hostSections.voice", icon: Mic }, { id: "workspaces", labelKey: "settings.hostSections.workspaces", icon: FolderGit2 }, { id: "providers", labelKey: "settings.hostSections.providers", icon: Boxes }, { id: "usage", labelKey: "settings.hostSections.usage", icon: Gauge }, @@ -256,6 +259,8 @@ function renderHostSettingsContent( return ; case "metadata": return ; + case "voice": + return ; case "workspaces": return ; case "providers": diff --git a/packages/app/src/screens/settings/voice/use-local-speech-models.ts b/packages/app/src/screens/settings/voice/use-local-speech-models.ts new file mode 100644 index 0000000000..b68a7b8135 --- /dev/null +++ b/packages/app/src/screens/settings/voice/use-local-speech-models.ts @@ -0,0 +1,71 @@ +import { useCallback } from "react"; +import { useTranslation } from "react-i18next"; +import type { LocalSpeechModel } from "@getpaseo/protocol/speech/rpc-schemas"; +import { useFetchQuery } from "@/data/query"; +import { useHostRuntimeClient, useHostRuntimeIsConnected } from "@/runtime/host-runtime"; +import { isLocalModelBusy } from "./voice-settings"; + +const DOWNLOAD_POLL_INTERVAL_MS = 1000; + +const localSpeechModelsQueryKey = (serverId: string) => ["speech-local-models", serverId] as const; + +export interface LocalSpeechModels { + models: LocalSpeechModel[]; + isLoading: boolean; + error: Error | null; + refresh: () => void; + download: (modelId: string) => Promise; + remove: (modelId: string) => Promise; +} + +// The daemon reports download progress through the list, so the list polls while any +// model is downloading or extracting. +export function useLocalSpeechModels(serverId: string): LocalSpeechModels { + const { t } = useTranslation(); + const client = useHostRuntimeClient(serverId); + const isConnected = useHostRuntimeIsConnected(serverId); + const query = useFetchQuery({ + queryKey: localSpeechModelsQueryKey(serverId), + enabled: Boolean(client && isConnected), + dataShape: "list", + staleTimeMs: 0, + queryFn: () => { + if (!client) throw new Error(t("workspace.terminal.hostDisconnected")); + return client.listLocalSpeechModels(); + }, + refetchInterval: (current) => + current.state.data?.some(isLocalModelBusy) ? DOWNLOAD_POLL_INTERVAL_MS : false, + }); + const { refetch } = query; + + const refresh = useCallback(() => { + void refetch(); + }, [refetch]); + + const download = useCallback( + async (modelId: string) => { + if (!client) return; + await client.downloadLocalSpeechModel(modelId); + await refetch(); + }, + [client, refetch], + ); + + const remove = useCallback( + async (modelId: string) => { + if (!client) return; + await client.deleteLocalSpeechModel(modelId); + await refetch(); + }, + [client, refetch], + ); + + return { + models: query.data ?? [], + isLoading: query.isLoading, + error: query.error, + refresh, + download, + remove, + }; +} diff --git a/packages/app/src/screens/settings/voice/voice-page.tsx b/packages/app/src/screens/settings/voice/voice-page.tsx new file mode 100644 index 0000000000..553ac017f7 --- /dev/null +++ b/packages/app/src/screens/settings/voice/voice-page.tsx @@ -0,0 +1,703 @@ +import { useCallback, useMemo, useRef, useState } from "react"; +import { Alert as NativeAlert, Text, View } from "react-native"; +import { useTranslation } from "react-i18next"; +import type { TFunction } from "i18next"; +import { StyleSheet } from "react-native-unistyles"; +import { Download, Languages, Trash2 } from "lucide-react-native"; +import type { + LocalSpeechModel, + SpeechSettings, + SpeechSettingsPatch, +} from "@getpaseo/protocol/speech/rpc-schemas"; +import { + SettingsCard, + SettingsCollapsibleRow, + SettingsInfoTip, + SettingsRow, + SettingsSection, + SettingsSelect, +} from "@/components/settings"; +import { Alert } from "@/components/ui/alert"; +import { Button } from "@/components/ui/button"; +import { FormTextInput } from "@/components/ui/form-field"; +import { LoadingSpinner } from "@/components/ui/loading-spinner"; +import { StatusBadge } from "@/components/ui/status-badge"; +import { Switch } from "@/components/ui/switch"; +import type { EditingTextInputHandle } from "@/components/ui/text-input"; +import { useDaemonConfig } from "@/hooks/use-daemon-config"; +import { useHostFeature } from "@/runtime/host-features"; +import { useHostRuntimeIsConnected } from "@/runtime/host-runtime"; +import { confirmDialog } from "@/utils/confirm-dialog"; +import { useLocalSpeechModels, type LocalSpeechModels } from "./use-local-speech-models"; +import { + SPEECH_PROVIDERS, + buildModelOptions, + buildOpenAiVoiceOptions, + formatLanguages, + formatModelSize, + getLocalModelState, + resolveSelectedModel, + resolveSelectedVoice, + summarizeLocalModels, + toSpeechProvider, + usesOpenAi, + type SpeechModelKind, + type SpeechProvider, +} from "./voice-settings"; + +type SavePatch = (patch: SpeechSettingsPatch) => void; + +function errorMessage(error: unknown): string { + return error instanceof Error ? error.message : String(error); +} + +export function HostVoicePage({ serverId }: { serverId: string }) { + const { t } = useTranslation(); + const isConnected = useHostRuntimeIsConnected(serverId); + const supported = useHostFeature(serverId, "speechSettings"); + const { config, patchConfig } = useDaemonConfig(serverId); + const localModels = useLocalSpeechModels(serverId); + const { refresh: refreshModels } = localModels; + + const save = useCallback( + (patch) => { + patchConfig({ speech: patch }) + .then(refreshModels) + .catch((error: unknown) => { + NativeAlert.alert(t("settings.voice.saveError"), errorMessage(error)); + }); + }, + [patchConfig, refreshModels, t], + ); + + if (!isConnected) { + return ; + } + if (!supported) { + return ; + } + const speech = config?.speech; + if (!speech) { + return ( + + + + ); + } + + return ( + + + + + + ); +} + +// --------------------------------------------------------------------------- +// Feature sections +// --------------------------------------------------------------------------- + +interface FeatureSectionProps { + speech: SpeechSettings; + localModels: LocalSpeechModel[]; + save: SavePatch; +} + +function DictationSection({ speech, localModels, save }: FeatureSectionProps) { + const { t } = useTranslation(); + const { dictation } = speech; + const setEnabled = useCallback((enabled: boolean) => save({ dictation: { enabled } }), [save]); + const setProvider = useCallback( + (provider: SpeechProvider) => save({ dictation: { stt: { provider } } }), + [save], + ); + const setModel = useCallback((model: string) => save({ dictation: { stt: { model } } }), [save]); + + return ( + + + + + + {dictation.enabled ? ( + + ) : null} + {dictation.enabled ? ( + + ) : null} + + + ); +} + +function VoiceModeSection({ speech, localModels, save }: FeatureSectionProps) { + const { t } = useTranslation(); + const { voiceMode } = speech; + const sttProvider = toSpeechProvider(voiceMode.stt.provider); + const ttsProvider = toSpeechProvider(voiceMode.tts.provider); + const setEnabled = useCallback((enabled: boolean) => save({ voiceMode: { enabled } }), [save]); + const setSttProvider = useCallback( + (provider: SpeechProvider) => save({ voiceMode: { stt: { provider } } }), + [save], + ); + const setSttModel = useCallback( + (model: string) => save({ voiceMode: { stt: { model } } }), + [save], + ); + const setTtsProvider = useCallback( + (provider: SpeechProvider) => save({ voiceMode: { tts: { provider } } }), + [save], + ); + const setTtsModel = useCallback( + (model: string) => save({ voiceMode: { tts: { model } } }), + [save], + ); + const setTtsVoice = useCallback( + (voice: string) => save({ voiceMode: { tts: { voice } } }), + [save], + ); + const voiceOptions = useMemo( + () => buildOpenAiVoiceOptions(resolveSelectedVoice(voiceMode.tts.voice)), + [voiceMode.tts.voice], + ); + + return ( + + + + + + {voiceMode.enabled ? ( + + ) : null} + {voiceMode.enabled ? ( + + ) : null} + {voiceMode.enabled ? ( + + ) : null} + {voiceMode.enabled ? ( + + ) : null} + {voiceMode.enabled && ttsProvider === "openai" ? ( + + ) : null} + + + ); +} + +function ProviderRow({ + label, + value, + onChange, + testID, +}: { + label: string; + value: SpeechProvider; + onChange: (provider: SpeechProvider) => void; + testID: string; +}) { + const { t } = useTranslation(); + const options = useMemo( + () => + SPEECH_PROVIDERS.map((provider) => ({ + value: provider, + label: t(`settings.voice.providers.${provider}`), + })), + [t], + ); + return ( + + ); +} + +function ModelRow({ + kind, + provider, + model, + localModels, + onChange, +}: { + kind: SpeechModelKind; + provider: SpeechProvider; + model: string | undefined; + localModels: LocalSpeechModel[]; + onChange: (model: string) => void; +}) { + const { t } = useTranslation(); + const selected = resolveSelectedModel({ kind, provider, model, localModels }); + const options = useMemo( + () => + buildModelOptions({ + kind, + provider, + selected, + localModels, + notDownloadedLabel: (label) => t("settings.voice.notDownloadedOption", { label }), + }), + [kind, localModels, provider, selected, t], + ); + return ( + + ); +} + +// --------------------------------------------------------------------------- +// Providers: one collapsible row per provider, each with a one-line summary +// --------------------------------------------------------------------------- + +function ProvidersSection({ + speech, + localModels, + save, +}: { + speech: SpeechSettings; + localModels: LocalSpeechModels; + save: SavePatch; +}) { + const { t } = useTranslation(); + const missingKey = usesOpenAi(speech) && !speech.openai.apiKeyConfigured; + return ( + + {missingKey ? ( + + + + ) : null} + + + + + + + ); +} + +function LocalProviderRow({ + kind, + localModels, +}: { + kind: SpeechModelKind; + localModels: LocalSpeechModels; +}) { + const { t } = useTranslation(); + const { models, isLoading, error, refresh, download, remove } = localModels; + const kindModels = useMemo(() => models.filter((model) => model.kind === kind), [kind, models]); + const label = + kind === "stt" + ? t("settings.voice.providerSettings.localStt") + : t("settings.voice.providerSettings.localTts"); + + if (error && kindModels.length === 0) { + return ( + + + + + + ); + } + return ( + + {kindModels.map((model) => ( + + ))} + + ); +} + +function localSummaryLabel( + models: readonly LocalSpeechModel[], + kind: SpeechModelKind, + t: TFunction, +): string { + const summary = summarizeLocalModels(models, kind); + switch (summary.kind) { + case "downloading": + return summary.percent === null + ? t("settings.voice.models.downloading") + : t("settings.voice.models.downloadingPercent", { percent: summary.percent }); + case "extracting": + return t("settings.voice.models.extracting"); + case "downloaded": + return t("settings.voice.providerSettings.downloaded", { count: summary.count }); + } +} + +function OpenAiProviderRow({ speech, save }: { speech: SpeechSettings; save: SavePatch }) { + const { t } = useTranslation(); + const { apiKeyConfigured, baseUrl } = speech.openai; + const keyInput = useRef(null); + const [draftKey, setDraftKey] = useState(""); + const [draftBaseUrl, setDraftBaseUrl] = useState(null); + + const saveKey = useCallback(() => { + save({ openai: { apiKey: draftKey } }); + keyInput.current?.replaceText(""); + setDraftKey(""); + }, [draftKey, save]); + const removeKey = useCallback(() => { + void confirmDialog({ + title: t("settings.voice.openai.removeKeyTitle"), + message: t("settings.voice.openai.removeKeyMessage"), + confirmLabel: t("settings.voice.openai.removeKey"), + destructive: true, + }).then((confirmed) => { + if (confirmed) save({ openai: { apiKey: null } }); + return undefined; + }); + }, [save, t]); + const saveBaseUrl = useCallback(() => { + if (draftBaseUrl === null) return; + save({ openai: { baseUrl: draftBaseUrl.trim() || null } }); + setDraftBaseUrl(null); + }, [draftBaseUrl, save]); + + const apiKeyInfo = useMemo( + () => ( + + ), + [t], + ); + const baseUrlInfo = useMemo( + () => ( + + ), + [t], + ); + + return ( + + + + + + {apiKeyConfigured ? ( + + ) : null} + + + + + + + + + + ); +} + +function LocalModelRow({ + model, + onDownload, + onDelete, +}: { + model: LocalSpeechModel; + onDownload: (modelId: string) => Promise; + onDelete: (modelId: string) => Promise; +}) { + const { t, i18n } = useTranslation(); + const state = getLocalModelState(model); + const [actionError, setActionError] = useState(null); + + const download = useCallback(() => { + setActionError(null); + onDownload(model.id).catch((error: unknown) => setActionError(errorMessage(error))); + }, [model.id, onDownload]); + const remove = useCallback(() => { + void confirmDialog({ + title: t("settings.voice.models.deleteTitle", { label: model.label }), + message: t("settings.voice.models.deleteMessage", { + size: formatModelSize(model.downloadSizeBytes), + }), + confirmLabel: t("settings.voice.models.delete"), + destructive: true, + }) + .then((confirmed) => { + if (!confirmed) return undefined; + setActionError(null); + return onDelete(model.id); + }) + .catch((error: unknown) => setActionError(errorMessage(error))); + }, [model.downloadSizeBytes, model.id, model.label, onDelete, t]); + + const languages = formatLanguages({ + languages: model.languages, + locale: i18n.language, + formatMore: (listed, count) => + t("settings.voice.models.moreLanguages", { languages: listed, count }), + }); + const details = useMemo( + () => ( + + + + {formatModelSize(model.downloadSizeBytes)} + {state.kind === "installed" ? ( + + ) : null} + {model.isDefault ? ( + + ) : null} + + ), + [ + languages, + model.description, + model.downloadSizeBytes, + model.id, + model.inUse, + model.isDefault, + model.label, + state.kind, + t, + ], + ); + + let control = null; + if (state.kind === "downloading") { + control = ( + + + + {state.percent === null + ? t("settings.voice.models.downloading") + : t("settings.voice.models.downloadingPercent", { percent: state.percent })} + + + ); + } else if (state.kind === "extracting") { + control = ( + + + {t("settings.voice.models.extracting")} + + ); + } else if (state.kind === "installed") { + control = model.inUse ? null : ( + + ); + } else { + control = ( + + ); + } + + return ( + + {control} + + ); +} + +const styles = StyleSheet.create((theme) => ({ + loading: { + paddingVertical: theme.spacing[8], + alignItems: "center", + }, + spinnerColor: { + color: theme.colors.foregroundMuted, + }, + sectionAlert: { + marginBottom: theme.spacing[3], + }, + inputControls: { + flexDirection: "row", + alignItems: "center", + flexWrap: "wrap", + gap: theme.spacing[2], + }, + input: { + minWidth: 220, + }, + details: { + flexDirection: "row", + alignItems: "center", + gap: theme.spacing[2], + }, + modelSize: { + color: theme.colors.foregroundMuted, + fontSize: theme.fontSize.sm, + fontVariant: ["tabular-nums"], + }, + progress: { + flexDirection: "row", + alignItems: "center", + gap: theme.spacing[2], + }, + progressText: { + color: theme.colors.foregroundMuted, + fontSize: theme.fontSize.sm, + fontVariant: ["tabular-nums"], + }, +})); diff --git a/packages/app/src/screens/settings/voice/voice-settings.test.ts b/packages/app/src/screens/settings/voice/voice-settings.test.ts new file mode 100644 index 0000000000..3dc828adbe --- /dev/null +++ b/packages/app/src/screens/settings/voice/voice-settings.test.ts @@ -0,0 +1,126 @@ +import { describe, expect, it } from "vitest"; +import type { LocalSpeechModel } from "@getpaseo/protocol/speech/rpc-schemas"; +import { + buildModelOptions, + formatLanguages, + formatModelSize, + getLocalModelState, + resolveSelectedModel, + summarizeLocalModels, +} from "./voice-settings"; + +function model(overrides: Partial): LocalSpeechModel { + return { + id: "parakeet-tdt-0.6b-v2-int8", + kind: "stt", + label: "Parakeet v2", + description: "Accurate English transcription.", + languages: ["en"], + downloadSizeBytes: 482_468_385, + isDefault: true, + inUse: true, + status: "installed", + ...overrides, + }; +} + +const catalog = [ + model({}), + model({ + id: "parakeet-tdt-110m-en-int8", + label: "Parakeet Small", + isDefault: false, + status: "not_installed", + inUse: false, + }), + model({ id: "kokoro-en-v0_19", kind: "tts", label: "Kokoro", inUse: false }), +]; + +describe("voice settings", () => { + it("falls back to the catalog default when no local model is saved", () => { + expect( + resolveSelectedModel({ + kind: "stt", + provider: "local", + model: undefined, + localModels: catalog, + }), + ).toBe("parakeet-tdt-0.6b-v2-int8"); + expect( + resolveSelectedModel({ + kind: "tts", + provider: "openai", + model: undefined, + localModels: catalog, + }), + ).toBe("tts-1"); + }); + + it("marks local models that still need a download", () => { + expect( + buildModelOptions({ + kind: "stt", + provider: "local", + selected: "parakeet-tdt-0.6b-v2-int8", + localModels: catalog, + notDownloadedLabel: (label) => `${label} (not downloaded)`, + }), + ).toEqual([ + { value: "parakeet-tdt-0.6b-v2-int8", label: "Parakeet v2" }, + { value: "parakeet-tdt-110m-en-int8", label: "Parakeet Small (not downloaded)" }, + ]); + }); + + it("keeps a custom OpenAI-compatible model id selectable", () => { + const options = buildModelOptions({ + kind: "stt", + provider: "openai", + selected: "my-whisper", + localModels: catalog, + notDownloadedLabel: (label) => label, + }); + expect(options.at(-1)).toEqual({ value: "my-whisper", label: "my-whisper" }); + }); + + it("reports download progress as a percentage", () => { + expect( + getLocalModelState(model({ status: "downloading", receivedBytes: 241, totalBytes: 482 })), + ).toEqual({ kind: "downloading", percent: 50 }); + expect(getLocalModelState(model({ status: "downloading" }))).toEqual({ + kind: "downloading", + percent: null, + }); + }); + + it("summarizes each local provider by its downloads", () => { + expect(summarizeLocalModels(catalog, "stt")).toEqual({ kind: "downloaded", count: 1 }); + expect(summarizeLocalModels(catalog, "tts")).toEqual({ kind: "downloaded", count: 1 }); + const downloading = [ + ...catalog, + model({ + id: "parakeet-tdt-0.6b-v3-int8", + status: "downloading", + receivedBytes: 1, + totalBytes: 4, + }), + ]; + expect(summarizeLocalModels(downloading, "stt")).toEqual({ kind: "downloading", percent: 25 }); + expect(summarizeLocalModels(downloading, "tts")).toEqual({ kind: "downloaded", count: 1 }); + }); + + it("formats download sizes", () => { + expect(formatModelSize(482_468_385)).toBe("482 MB"); + expect(formatModelSize(1_250_000_000)).toBe("1.3 GB"); + }); + + it("names a short language list and counts the rest of a long one", () => { + const formatMore = (languages: string, count: number) => `${languages} and ${count} more`; + expect(formatLanguages({ languages: ["en", "de"], locale: "en", formatMore })).toBe( + "English, German", + ); + const whisper = ["en", "zh", "de", "es", "ru", "ko", "fr", "ja", "pt", "tr", "pl", "ca"]; + expect(formatLanguages({ languages: whisper, locale: "en", formatMore })).toBe( + "English, Chinese, German, Spanish, Russian, Korean, French, Japanese, Portuguese, Turkish and 2 more", + ); + }); +}); diff --git a/packages/app/src/screens/settings/voice/voice-settings.ts b/packages/app/src/screens/settings/voice/voice-settings.ts new file mode 100644 index 0000000000..9e45616acc --- /dev/null +++ b/packages/app/src/screens/settings/voice/voice-settings.ts @@ -0,0 +1,173 @@ +import type { LocalSpeechModel, SpeechSettings } from "@getpaseo/protocol/speech/rpc-schemas"; + +// What the Voice page needs to know about speech providers and models, kept free of React +// so the option lists and model states are testable on their own. + +export type SpeechProvider = "local" | "openai"; +export type SpeechModelKind = "stt" | "tts"; + +export const SPEECH_PROVIDERS: readonly SpeechProvider[] = ["local", "openai"]; + +// OpenAI accepts free-form ids so self-hosted servers can use their own; these are the +// ones the page offers. A saved id outside the list is still shown as an option. +const OPENAI_MODELS: Record = { + stt: { + ids: ["whisper-1", "gpt-4o-transcribe", "gpt-4o-mini-transcribe"], + defaultId: "whisper-1", + }, + tts: { ids: ["tts-1", "tts-1-hd", "gpt-4o-mini-tts"], defaultId: "tts-1" }, +}; + +const OPENAI_VOICES = ["alloy", "ash", "coral", "echo", "fable", "nova", "onyx", "sage", "shimmer"]; +const DEFAULT_OPENAI_VOICE = "alloy"; + +export interface SpeechModelOption { + value: string; + label: string; +} + +export function toSpeechProvider(provider: string): SpeechProvider { + return provider === "openai" ? "openai" : "local"; +} + +export function resolveSelectedModel(input: { + kind: SpeechModelKind; + provider: SpeechProvider; + model: string | undefined; + localModels: readonly LocalSpeechModel[]; +}): string { + if (input.model) return input.model; + if (input.provider === "openai") return OPENAI_MODELS[input.kind].defaultId; + return input.localModels.find((m) => m.kind === input.kind && m.isDefault)?.id ?? ""; +} + +export function buildModelOptions(input: { + kind: SpeechModelKind; + provider: SpeechProvider; + selected: string; + localModels: readonly LocalSpeechModel[]; + notDownloadedLabel: (label: string) => string; +}): SpeechModelOption[] { + if (input.provider === "openai") { + return withSelected( + OPENAI_MODELS[input.kind].ids.map((id) => ({ value: id, label: id })), + input.selected, + ); + } + return input.localModels + .filter((model) => model.kind === input.kind) + .map((model) => ({ + value: model.id, + label: model.status === "installed" ? model.label : input.notDownloadedLabel(model.label), + })); +} + +export function resolveSelectedVoice(voice: string | undefined): string { + return voice ?? DEFAULT_OPENAI_VOICE; +} + +export function buildOpenAiVoiceOptions(selected: string): SpeechModelOption[] { + return withSelected( + OPENAI_VOICES.map((voice) => ({ value: voice, label: voice })), + selected, + ); +} + +function withSelected(options: SpeechModelOption[], selected: string): SpeechModelOption[] { + if (!selected || options.some((option) => option.value === selected)) return options; + return [...options, { value: selected, label: selected }]; +} + +export function usesOpenAi(settings: SpeechSettings): boolean { + return ( + (settings.dictation.enabled && settings.dictation.stt.provider === "openai") || + (settings.voiceMode.enabled && + (settings.voiceMode.stt.provider === "openai" || + settings.voiceMode.tts.provider === "openai")) + ); +} + +export type LocalModelState = + | { kind: "installed" } + | { kind: "not_installed" } + | { kind: "downloading"; percent: number | null } + | { kind: "extracting" } + | { kind: "failed"; error: string }; + +export function getLocalModelState(model: LocalSpeechModel): LocalModelState { + switch (model.status) { + case "installed": + return { kind: "installed" }; + case "downloading": { + const total = model.totalBytes ?? 0; + const received = model.receivedBytes ?? 0; + return { + kind: "downloading", + percent: total > 0 ? Math.min(100, Math.floor((received / total) * 100)) : null, + }; + } + case "extracting": + return { kind: "extracting" }; + case "failed": + return { kind: "failed", error: model.error ?? "" }; + default: + return { kind: "not_installed" }; + } +} + +export function isLocalModelBusy(model: LocalSpeechModel): boolean { + return model.status === "downloading" || model.status === "extracting"; +} + +// What a collapsed local provider row says: an active download wins over the count. +export type LocalModelsSummary = + | { kind: "downloading"; percent: number | null } + | { kind: "extracting" } + | { kind: "downloaded"; count: number }; + +export function summarizeLocalModels( + models: readonly LocalSpeechModel[], + kind: SpeechModelKind, +): LocalModelsSummary { + const kindModels = models.filter((model) => model.kind === kind); + const busy = kindModels.find(isLocalModelBusy); + if (busy) { + const state = getLocalModelState(busy); + if (state.kind === "downloading") return state; + return { kind: "extracting" }; + } + return { + kind: "downloaded", + count: kindModels.filter((model) => model.status === "installed").length, + }; +} + +export function formatModelSize(bytes: number): string { + const megabytes = bytes / 1_000_000; + if (megabytes >= 1000) return `${(megabytes / 1000).toFixed(1)} GB`; + return `${Math.round(megabytes)} MB`; +} + +// Multilingual models list close to a hundred languages; past this many the tooltip names +// the first ones and counts the rest. +const MAX_LISTED_LANGUAGES = 10; + +export function formatLanguages(input: { + languages: readonly string[]; + locale: string; + formatMore: (languages: string, count: number) => string; +}): string { + const names = languageNames(input.locale); + const listed = input.languages + .slice(0, MAX_LISTED_LANGUAGES) + .map((code) => names?.of(code) ?? code.toUpperCase()) + .join(", "); + const hidden = input.languages.length - MAX_LISTED_LANGUAGES; + return hidden > 0 ? input.formatMore(listed, hidden) : listed; +} + +// Hermes ships without Intl.DisplayNames; the language code is the fallback label there. +function languageNames(locale: string): Intl.DisplayNames | null { + if (typeof Intl.DisplayNames !== "function") return null; + return new Intl.DisplayNames([locale], { type: "language" }); +} diff --git a/packages/app/src/utils/host-routes.ts b/packages/app/src/utils/host-routes.ts index c13d6f25f9..c422c5ce6b 100644 --- a/packages/app/src/utils/host-routes.ts +++ b/packages/app/src/utils/host-routes.ts @@ -517,6 +517,7 @@ export const HOST_SECTION_SLUGS = [ "pair-device", "agents", "metadata", + "voice", "workspaces", "providers", "usage", diff --git a/packages/app/vitest.config.ts b/packages/app/vitest.config.ts index e2d67aca70..cfafa0da93 100644 --- a/packages/app/vitest.config.ts +++ b/packages/app/vitest.config.ts @@ -91,6 +91,8 @@ export default defineConfig({ "react-native-keyboard-controller", ], }, + // babel-preset-expo compiles JSX with the automatic runtime, so app modules don't import React. + esbuild: { jsx: "automatic" }, // The globals a React Native bundler defines, which esbuild is no longer there to supply for // the package excluded above. define: { diff --git a/packages/client/src/daemon-client.ts b/packages/client/src/daemon-client.ts index e0e4d9ddc5..384c4e391b 100644 --- a/packages/client/src/daemon-client.ts +++ b/packages/client/src/daemon-client.ts @@ -12,6 +12,7 @@ import { CreationClient } from "./creation/index.js"; import type { CreationSnapshot } from "@getpaseo/protocol/messages"; import type { z } from "zod"; import type { SessionEventSubscription } from "@getpaseo/protocol/messages"; +import type { LocalSpeechModel } from "@getpaseo/protocol/speech/rpc-schemas"; import type { ClientCapability } from "@getpaseo/protocol/client-capabilities"; import type { AgentAttentionNotificationPayload } from "@getpaseo/protocol/agent-attention-notification"; import { MAX_WEBSOCKET_MESSAGE_BYTES } from "@getpaseo/protocol/websocket-message-limit"; @@ -5264,6 +5265,34 @@ export class DaemonClient { }); } + async listLocalSpeechModels(requestId?: string): Promise { + const payload = await this.sendCorrelatedSessionRequest({ + requestId, + message: { type: "speech.local_models.list.request" }, + responseType: "speech.local_models.list.response", + }); + if (payload.error) throw new Error(payload.error); + return payload.models; + } + + async downloadLocalSpeechModel(modelId: string, requestId?: string): Promise { + const payload = await this.sendCorrelatedSessionRequest({ + requestId, + message: { type: "speech.local_models.download.request", modelId }, + responseType: "speech.local_models.download.response", + }); + if (payload.error) throw new Error(payload.error); + } + + async deleteLocalSpeechModel(modelId: string, requestId?: string): Promise { + const payload = await this.sendCorrelatedSessionRequest({ + requestId, + message: { type: "speech.local_models.delete.request", modelId }, + responseType: "speech.local_models.delete.response", + }); + if (payload.error) throw new Error(payload.error); + } + sendBrowserAutomationExecuteResponse(response: BrowserAutomationExecuteResponse): void { this.sendSessionMessageStrict(response); } diff --git a/packages/protocol/src/messages.ts b/packages/protocol/src/messages.ts index a7a2182d94..9c944d74f2 100644 --- a/packages/protocol/src/messages.ts +++ b/packages/protocol/src/messages.ts @@ -63,6 +63,16 @@ import { ScheduleRunOnceResponseSchema, ScheduleUpdateResponseSchema, } from "./schedule/rpc-schemas.js"; +import { + SpeechLocalModelsDeleteRequestSchema, + SpeechLocalModelsDeleteResponseSchema, + SpeechLocalModelsDownloadRequestSchema, + SpeechLocalModelsDownloadResponseSchema, + SpeechLocalModelsListRequestSchema, + SpeechLocalModelsListResponseSchema, + SpeechSettingsPatchSchema, + SpeechSettingsSchema, +} from "./speech/rpc-schemas.js"; import { LoopRunRequestSchema, LoopListRequestSchema, @@ -208,6 +218,8 @@ export const MutableDaemonConfigSchema = z skills: z.object({ selection: AgentSkillSelectionSchema.optional() }).strict().optional(), pluginsEnabled: z.boolean().optional(), plugins: z.record(PluginIdSchema, PluginSourceSchema).optional(), + // COMPAT(speechSettings): added in v0.12.0, optional because older daemons omit it. + speech: SpeechSettingsSchema.optional(), }) .passthrough(); @@ -228,6 +240,7 @@ export const MutableDaemonConfigPatchSchema = z agentProfiles: z.array(AgentProfileSchema).optional(), pluginsEnabled: z.boolean().optional(), plugins: z.record(PluginIdSchema, PluginSourceSchema).optional(), + speech: SpeechSettingsPatchSchema.optional(), }) .partial() .passthrough(); @@ -3390,6 +3403,9 @@ export const SessionInboundMessageSchema = z.discriminatedUnion("type", [ ChatWaitRequestSchema, ScheduleCreateRequestSchema, ScheduleListRequestSchema, + SpeechLocalModelsListRequestSchema, + SpeechLocalModelsDownloadRequestSchema, + SpeechLocalModelsDeleteRequestSchema, ScheduleInspectRequestSchema, ScheduleLogsRequestSchema, SchedulePauseRequestSchema, @@ -3620,6 +3636,8 @@ export const ServerInfoStatusPayloadSchema = z daemonStatusRpc: z.boolean().optional(), // COMPAT(daemonConfigReload): added in v0.4.0, remove gate after 2027-02-14. daemonConfigReload: z.boolean().optional(), + // COMPAT(speechSettings): added in v0.12.0, remove gate after 2027-04-18. + speechSettings: z.boolean().optional(), // COMPAT(relayConfig): added in v0.2.6, remove gate after 2027-01-31. relayConfig: z.boolean().optional(), // COMPAT(pushTokenRevocation): added in v0.3.2, remove gate after 2027-02-10. @@ -7061,6 +7079,9 @@ export const SessionOutboundMessageSchema = z.discriminatedUnion("type", [ ChatWaitResponseSchema, ScheduleCreateResponseSchema, ScheduleListResponseSchema, + SpeechLocalModelsListResponseSchema, + SpeechLocalModelsDownloadResponseSchema, + SpeechLocalModelsDeleteResponseSchema, ScheduleInspectResponseSchema, ScheduleLogsResponseSchema, SchedulePauseResponseSchema, diff --git a/packages/protocol/src/speech/rpc-schemas.ts b/packages/protocol/src/speech/rpc-schemas.ts new file mode 100644 index 0000000000..9e8dec9749 --- /dev/null +++ b/packages/protocol/src/speech/rpc-schemas.ts @@ -0,0 +1,165 @@ +import { z } from "zod"; + +// --------------------------------------------------------------------------- +// Speech settings: the `speech` section of the mutable daemon config. +// --------------------------------------------------------------------------- + +// Provider ids stay plain strings so an old client can parse a daemon that adds one. +const SpeechSttSettingsSchema = z + .object({ + provider: z.string(), + // Omitted means the provider's default model. + model: z.string().optional(), + language: z.string().optional(), + }) + .passthrough(); + +const SpeechTtsSettingsSchema = z + .object({ + provider: z.string(), + model: z.string().optional(), + voice: z.string().optional(), + }) + .passthrough(); + +export const SpeechSettingsSchema = z + .object({ + dictation: z + .object({ + enabled: z.boolean(), + stt: SpeechSttSettingsSchema, + }) + .passthrough(), + voiceMode: z + .object({ + enabled: z.boolean(), + stt: SpeechSttSettingsSchema, + tts: SpeechTtsSettingsSchema, + }) + .passthrough(), + openai: z + .object({ + // The key itself is write-only; readers only learn whether one is set. + apiKeyConfigured: z.boolean(), + baseUrl: z.string().optional(), + }) + .passthrough(), + }) + .passthrough(); + +// null clears a field back to its default. Changing `provider` without `model` clears +// the model, because model ids belong to one provider. +const SpeechSttSettingsPatchSchema = z + .object({ + provider: z.string().optional(), + model: z.string().nullable().optional(), + language: z.string().nullable().optional(), + }) + .passthrough(); + +const SpeechTtsSettingsPatchSchema = z + .object({ + provider: z.string().optional(), + model: z.string().nullable().optional(), + voice: z.string().nullable().optional(), + }) + .passthrough(); + +export const SpeechSettingsPatchSchema = z + .object({ + dictation: z + .object({ + enabled: z.boolean().optional(), + stt: SpeechSttSettingsPatchSchema.optional(), + }) + .passthrough() + .optional(), + voiceMode: z + .object({ + enabled: z.boolean().optional(), + stt: SpeechSttSettingsPatchSchema.optional(), + tts: SpeechTtsSettingsPatchSchema.optional(), + }) + .passthrough() + .optional(), + openai: z + .object({ + apiKey: z.string().nullable().optional(), + baseUrl: z.string().nullable().optional(), + }) + .passthrough() + .optional(), + }) + .passthrough(); + +export type SpeechSettings = z.infer; +export type SpeechSettingsPatch = z.infer; + +// --------------------------------------------------------------------------- +// Local speech models: catalog, install state, download, delete. +// --------------------------------------------------------------------------- + +export const LocalSpeechModelSchema = z + .object({ + id: z.string(), + // "stt" or "tts". + kind: z.string(), + label: z.string(), + description: z.string(), + languages: z.array(z.string()), + downloadSizeBytes: z.number(), + isDefault: z.boolean(), + // Selected by an enabled feature; the daemon keeps it installed. + inUse: z.boolean(), + // "installed" | "not_installed" | "downloading" | "extracting" | "failed". + status: z.string(), + receivedBytes: z.number().optional(), + totalBytes: z.number().optional(), + error: z.string().optional(), + }) + .passthrough(); + +export type LocalSpeechModel = z.infer; + +export const SpeechLocalModelsListRequestSchema = z.object({ + type: z.literal("speech.local_models.list.request"), + requestId: z.string(), +}); + +export const SpeechLocalModelsDownloadRequestSchema = z.object({ + type: z.literal("speech.local_models.download.request"), + requestId: z.string(), + modelId: z.string(), +}); + +export const SpeechLocalModelsDeleteRequestSchema = z.object({ + type: z.literal("speech.local_models.delete.request"), + requestId: z.string(), + modelId: z.string(), +}); + +export const SpeechLocalModelsListResponseSchema = z.object({ + type: z.literal("speech.local_models.list.response"), + payload: z.object({ + requestId: z.string(), + models: z.array(LocalSpeechModelSchema), + error: z.string().nullable(), + }), +}); + +// Resolves once the download starts; poll the list for progress. +export const SpeechLocalModelsDownloadResponseSchema = z.object({ + type: z.literal("speech.local_models.download.response"), + payload: z.object({ + requestId: z.string(), + error: z.string().nullable(), + }), +}); + +export const SpeechLocalModelsDeleteResponseSchema = z.object({ + type: z.literal("speech.local_models.delete.response"), + payload: z.object({ + requestId: z.string(), + error: z.string().nullable(), + }), +}); diff --git a/packages/server/src/server/authorization/operation-permissions.ts b/packages/server/src/server/authorization/operation-permissions.ts index e7e77bc0d3..bedd6f79c6 100644 --- a/packages/server/src/server/authorization/operation-permissions.ts +++ b/packages/server/src/server/authorization/operation-permissions.ts @@ -174,6 +174,9 @@ const INBOUND_PERMISSION = { set_agent_model_request: "workspace.write", set_agent_thinking_request: "workspace.write", set_daemon_config_request: "daemon.manage", + "speech.local_models.list.request": "daemon.read", + "speech.local_models.download.request": "daemon.manage", + "speech.local_models.delete.request": "daemon.manage", set_voice_mode: "workspace.write", shutdown_server_request: "daemon.manage", start_workspace_script_request: "workspace.write", @@ -399,6 +402,9 @@ const OUTBOUND_PERMISSION = { set_agent_model_response: "workspace.write", set_agent_thinking_response: "workspace.write", set_daemon_config_response: "daemon.manage", + "speech.local_models.list.response": "daemon.read", + "speech.local_models.download.response": "daemon.manage", + "speech.local_models.delete.response": "daemon.manage", set_voice_mode_response: "workspace.write", start_workspace_script_response: "workspace.write", stash_list_response: "workspace.read", diff --git a/packages/server/src/server/bootstrap.ts b/packages/server/src/server/bootstrap.ts index 727c7b7899..12bd803da9 100644 --- a/packages/server/src/server/bootstrap.ts +++ b/packages/server/src/server/bootstrap.ts @@ -131,6 +131,7 @@ import type { OpenAiSpeechProviderConfig } from "./speech/providers/openai/confi import type { LocalSpeechProviderConfig } from "./speech/providers/local/config.js"; import type { RequestedSpeechProviders } from "./speech/speech-types.js"; import { createSpeechService } from "./speech/speech-runtime.js"; +import { createSpeechSettingsOwner } from "./speech/speech-settings.js"; import { AgentManager } from "./agent/agent-manager.js"; import { AgentStorage } from "./agent/agent-storage.js"; import { attachAgentStoragePersistence } from "./persistence-hooks.js"; @@ -370,6 +371,8 @@ export interface PaseoSpeechSttLanguages { export interface PaseoSpeechConfig { providers: RequestedSpeechProviders; + // Where local models live, whether or not a feature currently uses the local provider. + localModelsDir: string; sttLanguages?: PaseoSpeechSttLanguages; local?: PaseoLocalSpeechConfig; } @@ -537,6 +540,10 @@ function resolveExpressTrustProxySetting(config: PaseoDaemonConfig): true | stri return config.trustedProxies ?? ["loopback"]; } +function resolveReloadEnv(config: PaseoDaemonConfig): NodeJS.ProcessEnv { + return config.configReload?.env ?? process.env; +} + function createInitialMutableDaemonConfig(config: PaseoDaemonConfig): MutableDaemonConfig { const providers = config.providerOverrides ?? {}; @@ -596,13 +603,24 @@ export async function createPaseoDaemon( const elapsed = () => `${(performance.now() - bootstrapStart).toFixed(0)}ms`; const daemonVersion = config.daemonVersion ?? resolveDaemonVersion(import.meta.url); const initialMutableConfig = createInitialMutableDaemonConfig(config); + const speechService = createSpeechService({ + logger, + openaiConfig: config.openai, + speechConfig: config.speech, + }); + logger.info({ elapsed: elapsed() }, "Speech service created"); const daemonConfigStore = new DaemonConfigStore(config.paseoHome, initialMutableConfig, logger, { relayEnabledMutable: config.relayEnabledMutable ?? true, startupPersisted: config.configReload?.startupPersisted, + speech: createSpeechSettingsOwner({ + paseoHome: config.paseoHome, + env: resolveReloadEnv(config), + speechService, + }), reloadSource: { resolve: (persisted) => { const reloaded = resolveConfigFromPersisted(config.paseoHome, persisted, { - env: config.configReload?.env ?? process.env, + env: resolveReloadEnv(config), cli: config.configReload?.cli, relayEnabledFallback: config.configReload?.relayEnabledFallback, }); @@ -1591,13 +1609,6 @@ export async function createPaseoDaemon( logger.info({ route: agentMcpRoute, enabled: mcpEnabled }, "Agent MCP route mounted"); } - const speechService = createSpeechService({ - logger, - openaiConfig: config.openai, - speechConfig: config.speech, - }); - logger.info({ elapsed: elapsed() }, "Speech service created"); - logger.info({ elapsed: elapsed() }, "Bootstrap complete, ready to start listening"); const start = async () => { diff --git a/packages/server/src/server/daemon-client.e2e.test.ts b/packages/server/src/server/daemon-client.e2e.test.ts index b6518d773d..a442dfe6a9 100644 --- a/packages/server/src/server/daemon-client.e2e.test.ts +++ b/packages/server/src/server/daemon-client.e2e.test.ts @@ -1318,9 +1318,9 @@ test("emits disabled voice capability reasons on fresh daemon startup", async () expect(voice).toBeTruthy(); expect(voice?.dictation.enabled).toBe(false); - expect(voice?.dictation.reason).toBe("Dictation is disabled in daemon config."); + expect(voice?.dictation.reason).toBe("Dictation is off. Turn it on in Settings > Voice."); expect(voice?.voice.enabled).toBe(false); - expect(voice?.voice.reason).toBe("Realtime voice is disabled in daemon config."); + expect(voice?.voice.reason).toBe("Voice mode is off. Turn it on in Settings > Voice."); } finally { await client.close().catch(() => undefined); await isolatedCtx.cleanup(); diff --git a/packages/server/src/server/daemon-config-store.test.ts b/packages/server/src/server/daemon-config-store.test.ts index 3629a0fd18..61ee2a1dd0 100644 --- a/packages/server/src/server/daemon-config-store.test.ts +++ b/packages/server/src/server/daemon-config-store.test.ts @@ -5,6 +5,7 @@ import { afterEach, describe, expect, test } from "vitest"; import { DaemonConfigStore, applyMutableProviderConfigToOverrides } from "./daemon-config-store.js"; import { loadPersistedConfig } from "./persisted-config.js"; +import { readSpeechSettings } from "./speech/speech-settings.js"; import type { PersistedConfig } from "./persisted-config.js"; import type { MutableDaemonConfig } from "@getpaseo/protocol/messages"; @@ -1193,3 +1194,110 @@ describe("DaemonConfigStore reload", () => { }); }); }); + +describe("DaemonConfigStore speech settings", () => { + const tempDirs: string[] = []; + + afterEach(() => { + for (const dir of tempDirs) rmSync(dir, { recursive: true, force: true }); + }); + + function createSpeechStore(initialPersisted?: PersistedConfig) { + const paseoHome = mkdtempSync(path.join(tmpdir(), "paseo-daemon-config-speech-")); + tempDirs.push(paseoHome); + if (initialPersisted) { + writeFileSync( + path.join(paseoHome, "config.json"), + `${JSON.stringify(initialPersisted, null, 2)}\n`, + ); + } + const persisted = loadPersistedConfig(paseoHome); + const applied: PersistedConfig[] = []; + const speech = { + read: (next: PersistedConfig) => readSpeechSettings(next, {}), + apply: (next: PersistedConfig) => applied.push(next), + }; + const store = new DaemonConfigStore(paseoHome, reloadableConfig(persisted), undefined, { + speech, + reloadSource: { + resolve: (next) => ({ mutable: reloadableConfig(next), overrideControlledPaths: [] }), + }, + }); + return { paseoHome, store, applied, persisted }; + } + + test("keeps voice mode off until it is turned on", () => { + const { paseoHome, store } = createSpeechStore(); + + expect(store.get().speech?.dictation.enabled).toBe(true); + expect(store.get().speech?.voiceMode.enabled).toBe(false); + + const result = store.patch({ speech: { voiceMode: { enabled: true } } }); + + expect(result.speech?.voiceMode.enabled).toBe(true); + expect(loadPersistedConfig(paseoHome).features?.voiceMode?.enabled).toBe(true); + }); + + test("stores the OpenAI key without ever returning it", () => { + const { paseoHome, store, applied } = createSpeechStore(); + + const result = store.patch({ + speech: { dictation: { stt: { provider: "openai" } }, openai: { apiKey: " sk-test " } }, + }); + + expect(result.speech?.openai).toEqual({ apiKeyConfigured: true }); + expect(JSON.stringify(result)).not.toContain("sk-test"); + expect(loadPersistedConfig(paseoHome).providers?.openai?.apiKey).toBe("sk-test"); + expect(applied).toHaveLength(1); + }); + + test("applies a replaced key even though the visible settings do not change", () => { + const { store, applied } = createSpeechStore({ + version: 1, + providers: { openai: { apiKey: "sk-old" } }, + }); + + store.patch({ speech: { openai: { apiKey: "sk-new" } } }); + + expect(applied.map((next) => next.providers?.openai?.apiKey)).toEqual(["sk-new"]); + }); + + test("switching provider clears the previous provider's model", () => { + const { paseoHome, store } = createSpeechStore({ + version: 1, + features: { dictation: { stt: { provider: "local", model: "parakeet-tdt-0.6b-v3-int8" } } }, + }); + + const result = store.patch({ speech: { dictation: { stt: { provider: "openai" } } } }); + + expect(result.speech?.dictation.stt).toEqual({ provider: "openai" }); + expect(loadPersistedConfig(paseoHome).features?.dictation?.stt).toEqual({ + provider: "openai", + }); + }); + + test("rejects a local model id outside the catalog without persisting it", () => { + const { paseoHome, store, applied } = createSpeechStore(); + + expect(() => + store.patch({ speech: { voiceMode: { tts: { provider: "local", model: "nope" } } } }), + ).toThrow("Unknown local voice model: nope"); + expect(loadPersistedConfig(paseoHome).features?.voiceMode).toBeUndefined(); + expect(applied).toEqual([]); + }); + + test("reload applies hand-edited speech settings without a restart", () => { + const { paseoHome, store, applied, persisted } = createSpeechStore(); + writeFileSync( + path.join(paseoHome, "config.json"), + `${JSON.stringify({ ...persisted, features: { voiceMode: { enabled: false } } }, null, 2)}\n`, + ); + + const result = store.reload(); + + expect(result.appliedPaths).toContain("features.voiceMode"); + expect(result.restartRequiredPaths).toEqual([]); + expect(store.get().speech?.voiceMode.enabled).toBe(false); + expect(applied).toHaveLength(1); + }); +}); diff --git a/packages/server/src/server/daemon-config-store.ts b/packages/server/src/server/daemon-config-store.ts index 3e285bded0..432c560a32 100644 --- a/packages/server/src/server/daemon-config-store.ts +++ b/packages/server/src/server/daemon-config-store.ts @@ -9,6 +9,8 @@ import { MutableDaemonConfigPatchSchema, } from "@getpaseo/protocol/messages"; import type { AgentSkillSelection } from "@getpaseo/protocol/messages"; +import type { SpeechSettings, SpeechSettingsPatch } from "@getpaseo/protocol/speech/rpc-schemas"; +import { mergeSpeechSettingsPatch, type SpeechSettingsOwner } from "./speech/speech-settings.js"; export type { MutableDaemonConfig, MutableDaemonConfigPatch } from "@getpaseo/protocol/messages"; @@ -31,8 +33,11 @@ interface SupportedMutableConfigPatch { skills?: MutableDaemonConfig["skills"]; pluginsEnabled?: boolean; plugins?: MutableDaemonConfig["plugins"]; + speech?: SpeechSettingsPatch; } +const SPEECH_PERSISTED_PATHS = ["features.dictation", "features.voiceMode", "providers.openai"]; + interface LoggerLike { child(bindings: Record): LoggerLike; info(...args: unknown[]): void; @@ -276,6 +281,7 @@ function pickSupportedPatchFields(patch: MutableDaemonConfigPatch): SupportedMut ...(patch.agentProfiles !== undefined ? { agentProfiles: patch.agentProfiles } : {}), ...(patch.pluginsEnabled !== undefined ? { pluginsEnabled: patch.pluginsEnabled } : {}), ...(patch.plugins !== undefined ? { plugins: patch.plugins } : {}), + ...(patch.speech !== undefined ? { speech: patch.speech } : {}), }; } @@ -318,6 +324,7 @@ export class DaemonConfigStore { private readonly relayEnabledMutable: boolean; private readonly reloadSource: DaemonConfigReloadSource | undefined; private readonly startupPersisted: PersistedConfig; + private readonly speechOwner: SpeechSettingsOwner | undefined; private lastKnownPersisted: PersistedConfig; constructor( @@ -328,18 +335,21 @@ export class DaemonConfigStore { relayEnabledMutable?: boolean; reloadSource?: DaemonConfigReloadSource; startupPersisted?: PersistedConfig; + speech?: SpeechSettingsOwner; } = {}, ) { this.paseoHome = paseoHome; this.logger = getLogger(logger); + this.relayEnabledMutable = options.relayEnabledMutable ?? true; + this.reloadSource = options.reloadSource; + this.speechOwner = options.speech; + this.startupPersisted = options.startupPersisted ?? loadPersistedConfig(paseoHome, this.logger); + this.lastKnownPersisted = this.startupPersisted; this.current = MutableDaemonConfigSchema.parse({ ...initial, relay: initial.relay ?? { enabled: true }, + ...this.readSpeech(this.startupPersisted), }); - this.relayEnabledMutable = options.relayEnabledMutable ?? true; - this.reloadSource = options.reloadSource; - this.startupPersisted = options.startupPersisted ?? loadPersistedConfig(paseoHome, this.logger); - this.lastKnownPersisted = this.startupPersisted; } public get(): MutableDaemonConfig { @@ -361,9 +371,11 @@ export class DaemonConfigStore { "Relay is controlled by a daemon launch override. Remove PASEO_RELAY_ENABLED or the relay CLI flag before changing it here.", ); } - const { removeProviders = [], ...configPatch } = parsedPatch; + const { removeProviders = [], ...patch } = parsedPatch; + const { speech: speechPatch, ...configPatch } = patch; const removedProviders = Array.from(new Set(removeProviders)); const merged = deepMerge(this.current, configPatch); + if (speechPatch !== undefined) merged.speech = this.readPatchedSpeech(speechPatch); if (parsedPatch.skills?.selection !== undefined) { merged.skills = { selection: parsedPatch.skills.selection }; } @@ -377,30 +389,47 @@ export class DaemonConfigStore { const configChanged = !isEqualValue(this.current, next); - if (!configChanged && removedProviders.length === 0) { + // A replaced API key leaves the speech view unchanged, so a speech patch always persists. + if (!configChanged && removedProviders.length === 0 && speechPatch === undefined) { return this.current; } const { previous: persistedBeforePatch, knownNext } = this.persistConfig( - configPatch, + patch, removedProviders, ); - if (!configChanged) { - this.lastKnownPersisted = knownNext; - return this.current; - } - - try { - this.applyReplacement(next, { removedProviders }); - this.lastKnownPersisted = knownNext; - } catch (error) { - savePersistedConfig(this.paseoHome, persistedBeforePatch, this.logger); - throw error; + const speechChanged = this.speechPersistedChanged(this.lastKnownPersisted, knownNext); + if (configChanged) { + try { + this.applyReplacement(next, { removedProviders }); + } catch (error) { + savePersistedConfig(this.paseoHome, persistedBeforePatch, this.logger); + throw error; + } } + this.lastKnownPersisted = knownNext; + if (speechChanged) this.speechOwner?.apply(knownNext); return this.current; } + private readSpeech(persisted: PersistedConfig): Pick { + return this.speechOwner ? { speech: this.speechOwner.read(persisted) } : {}; + } + + private readPatchedSpeech(speechPatch: SpeechSettingsPatch): SpeechSettings { + if (!this.speechOwner) { + throw new Error("Speech settings are unavailable for this daemon instance"); + } + return this.speechOwner.read(mergeSpeechSettingsPatch(this.lastKnownPersisted, speechPatch)); + } + + private speechPersistedChanged(previous: PersistedConfig, next: PersistedConfig): boolean { + return diffPaths(previous, next).some((path) => + SPEECH_PERSISTED_PATHS.some((owner) => pathBelongsTo(path, owner)), + ); + } + public reload(): DaemonConfigReloadResult { if (!this.reloadSource) { throw new Error("Daemon config reload is unavailable for this daemon instance"); @@ -413,6 +442,7 @@ export class DaemonConfigStore { const desired = MutableDaemonConfigSchema.parse({ ...resolved.mutable, plugins: this.current.plugins, + ...this.readSpeech(persisted), }); const changedSinceLastApply = diffPaths(this.lastKnownPersisted, persisted); const overrideControlledPaths = compactOwnedPaths( @@ -421,7 +451,7 @@ export class DaemonConfigStore { ), resolved.overrideControlledPaths, ); - const appliedPaths = RELOADABLE_PATHS.filter((persistedPath) => { + const appliedPaths: string[] = RELOADABLE_PATHS.filter((persistedPath) => { if (resolved.overrideControlledPaths.some((owner) => pathBelongsTo(persistedPath, owner))) { return false; } @@ -438,6 +468,8 @@ export class DaemonConfigStore { diffPaths(this.startupPersisted, persisted).filter((path) => { if (path === "$schema" || path === "version") return false; if (RELOADABLE_PATHS.some((owner) => pathBelongsTo(path, owner))) return false; + if (this.speechOwner && SPEECH_PERSISTED_PATHS.some((owner) => pathBelongsTo(path, owner))) + return false; return !resolved.overrideControlledPaths.some((owner) => pathBelongsTo(path, owner)); }), [], @@ -446,8 +478,17 @@ export class DaemonConfigStore { const removedProviders = Object.keys(this.current.providers).filter( (provider) => !(provider in desired.providers), ); + const speechChanged = this.speechPersistedChanged(this.lastKnownPersisted, persisted); this.applyReplacement(desired, { removedProviders }); this.lastKnownPersisted = persisted; + if (speechChanged && this.speechOwner) { + this.speechOwner.apply(persisted); + appliedPaths.push( + ...SPEECH_PERSISTED_PATHS.filter((owner) => + changedSinceLastApply.some((path) => pathBelongsTo(path, owner)), + ), + ); + } return { appliedPaths: [...appliedPaths].sort(), @@ -584,8 +625,9 @@ function mergeMutablePatchIntoPersistedConfig(params: { const { persisted, patch, removeProviders, persistRelayEnabled } = params; const daemon = mergeMutableDaemonPatch(persisted.daemon, patch, persistRelayEnabled); const agents = mergeMutableAgentPatch(persisted.agents, patch, removeProviders); + const withSpeech = patch.speech ? mergeSpeechSettingsPatch(persisted, patch.speech) : persisted; return { - ...persisted, + ...withSpeech, ...(patch.pluginsEnabled !== undefined ? { pluginsEnabled: patch.pluginsEnabled } : {}), ...(patch.plugins !== undefined ? { plugins: patch.plugins } : {}), ...(daemon ? { daemon } : { daemon: undefined }), diff --git a/packages/server/src/server/session.ts b/packages/server/src/server/session.ts index 08d3f0eefb..aea8a38dc9 100644 --- a/packages/server/src/server/session.ts +++ b/packages/server/src/server/session.ts @@ -182,6 +182,10 @@ import { import { ScheduleSession } from "./session/schedule/schedule-session.js"; import { ProviderCatalogSession } from "./session/provider/provider-catalog-session.js"; import { UsageSession } from "./session/usage/usage-session.js"; +import { + SpeechModelsSession, + type SpeechModelsSessionOptions, +} from "./session/speech/speech-models-session.js"; import { WorkspaceFilesSession } from "./session/files/workspace-files-session.js"; import { AgentConfigSession } from "./session/agent-config/agent-config-session.js"; import { ProjectConfigSession } from "./session/project-config/project-config-session.js"; @@ -525,6 +529,7 @@ export interface SessionOptions { stt: Resolvable; sttLanguage?: string; tts: Resolvable; + speechModels?: SpeechModelsSessionOptions["runtime"]; terminalManager: TerminalManager | null; providerSnapshotManager: ProviderSnapshotManager; hubExecutionAgents?: HubExecutionAgents; @@ -792,6 +797,7 @@ export class Session { private readonly scheduleSession: ScheduleSession; private readonly providerCatalogSession: ProviderCatalogSession; private readonly usageSession: UsageSession; + private readonly speechModelsSession: SpeechModelsSession; private readonly workspaceFilesSession: WorkspaceFilesSession; private readonly agentConfigSession: AgentConfigSession; private readonly projectConfigSession: ProjectConfigSession; @@ -840,6 +846,7 @@ export class Session { stt, sttLanguage, tts, + speechModels, terminalManager, providerSnapshotManager, serviceProxy, @@ -1005,6 +1012,10 @@ export class Session { runtime: pluginRuntime, logger: this.sessionLogger, }); + this.speechModelsSession = new SpeechModelsSession({ + emit: (msg) => this.emit(msg), + runtime: speechModels, + }); this.agentConfigSession = new AgentConfigSession({ host: { emit: (msg) => this.emit(msg), @@ -2336,12 +2347,19 @@ export class Session { this.dispatchOrchestrationSkillsMessage(msg) ?? this.dispatchPluginDirectoryMessage(msg) ?? this.dispatchPluginMessage(msg) ?? - this.dispatchTerminalMessage(msg) ?? - this.dispatchScheduleMessage(msg) ?? + this.dispatchHostServiceMessage(msg) ?? this.dispatchMiscMessage(msg); if (promise) await promise; } + private dispatchHostServiceMessage(msg: SessionInboundMessage): Promise | undefined { + return ( + this.dispatchTerminalMessage(msg) ?? + this.dispatchScheduleMessage(msg) ?? + this.speechModelsSession.dispatch(msg) + ); + } + private dispatchWorkspaceLifecycleMessage(msg: SessionInboundMessage): Promise | undefined { return ( this.dispatchWorkspaceStateMessage(msg) ?? diff --git a/packages/server/src/server/session/speech/speech-models-session.test.ts b/packages/server/src/server/session/speech/speech-models-session.test.ts new file mode 100644 index 0000000000..9624475ad4 --- /dev/null +++ b/packages/server/src/server/session/speech/speech-models-session.test.ts @@ -0,0 +1,106 @@ +import { describe, expect, test } from "vitest"; + +import type { SessionOutboundMessage } from "../../messages.js"; +import { + listLocalSpeechModels, + type LocalSpeechModelId, +} from "../../speech/providers/local/models.js"; +import { + LocalSpeechModelInUseError, + type LocalSpeechModelStatus, +} from "../../speech/speech-runtime.js"; +import { SpeechModelsSession } from "./speech-models-session.js"; + +const [firstModel] = listLocalSpeechModels(); + +function createSession(models: LocalSpeechModelStatus[], inUse: LocalSpeechModelId[] = []) { + const emitted: SessionOutboundMessage[] = []; + const downloads: LocalSpeechModelId[] = []; + const session = new SpeechModelsSession({ + emit: (message) => emitted.push(message), + runtime: { + listLocalModels: async () => models, + downloadLocalModel: (modelId) => { + downloads.push(modelId); + }, + deleteLocalModel: async (modelId) => { + if (inUse.includes(modelId)) throw new LocalSpeechModelInUseError(modelId); + }, + }, + }); + return { session, emitted, downloads }; +} + +describe("SpeechModelsSession", () => { + test("lists catalog details and download progress before anything is installed", async () => { + const { session, emitted } = createSession([ + { + spec: firstModel, + state: { status: "downloading", receivedBytes: 10, totalBytes: 40 }, + inUse: true, + }, + ]); + + await session.dispatch({ type: "speech.local_models.list.request", requestId: "r1" }); + + expect(emitted).toEqual([ + { + type: "speech.local_models.list.response", + payload: { + requestId: "r1", + error: null, + models: [ + { + id: firstModel.id, + kind: "stt", + label: firstModel.label, + description: firstModel.description, + languages: [...firstModel.languages], + downloadSizeBytes: firstModel.downloadSizeBytes, + isDefault: true, + inUse: true, + status: "downloading", + receivedBytes: 10, + totalBytes: 40, + }, + ], + }, + }, + ]); + }); + + test("rejects a model id outside the catalog without starting a download", () => { + const { session, emitted, downloads } = createSession([]); + + void session.dispatch({ + type: "speech.local_models.download.request", + requestId: "r2", + modelId: "../../etc", + }); + + expect(downloads).toEqual([]); + expect(emitted).toEqual([ + { + type: "speech.local_models.download.response", + payload: { requestId: "r2", error: "Unknown local speech model: ../../etc" }, + }, + ]); + }); + + test("reports why an in-use model cannot be deleted", async () => { + const { session, emitted } = createSession([], [firstModel.id]); + + await session.dispatch({ + type: "speech.local_models.delete.request", + requestId: "r3", + modelId: firstModel.id, + }); + + expect(emitted).toEqual([ + { + type: "speech.local_models.delete.response", + payload: { requestId: "r3", error: new LocalSpeechModelInUseError(firstModel.id).message }, + }, + ]); + }); +}); diff --git a/packages/server/src/server/session/speech/speech-models-session.ts b/packages/server/src/server/session/speech/speech-models-session.ts new file mode 100644 index 0000000000..d987623a8d --- /dev/null +++ b/packages/server/src/server/session/speech/speech-models-session.ts @@ -0,0 +1,124 @@ +import type { LocalSpeechModel } from "@getpaseo/protocol/speech/rpc-schemas"; + +import type { SessionInboundMessage, SessionOutboundMessage } from "../../messages.js"; +import { LocalSpeechModelIdSchema } from "../../speech/providers/local/models.js"; +import type { LocalSpeechModelStatus, SpeechService } from "../../speech/speech-runtime.js"; + +type SpeechModelsRuntime = Pick< + SpeechService, + "listLocalModels" | "downloadLocalModel" | "deleteLocalModel" +>; + +export interface SpeechModelsSessionOptions { + emit(message: SessionOutboundMessage): void; + // Absent when the daemon runs without speech. + runtime?: SpeechModelsRuntime | null; +} + +export class SpeechModelsSession { + constructor(private readonly options: SpeechModelsSessionOptions) {} + + // Handles the speech.local_models.* requests; returns undefined for any other message. + dispatch(msg: SessionInboundMessage): Promise | undefined { + switch (msg.type) { + case "speech.local_models.list.request": + return this.handleList(msg); + case "speech.local_models.download.request": + this.handleDownload(msg); + return Promise.resolve(); + case "speech.local_models.delete.request": + return this.handleDelete(msg); + default: + return undefined; + } + } + + private async handleList( + msg: Extract, + ): Promise { + try { + const models = await this.requireRuntime().listLocalModels(); + this.options.emit({ + type: "speech.local_models.list.response", + payload: { requestId: msg.requestId, models: models.map(toWireModel), error: null }, + }); + } catch (error) { + this.options.emit({ + type: "speech.local_models.list.response", + payload: { requestId: msg.requestId, models: [], error: describeError(error) }, + }); + } + } + + private handleDownload( + msg: Extract, + ): void { + try { + this.requireRuntime().downloadLocalModel(parseModelId(msg.modelId)); + this.options.emit({ + type: "speech.local_models.download.response", + payload: { requestId: msg.requestId, error: null }, + }); + } catch (error) { + this.options.emit({ + type: "speech.local_models.download.response", + payload: { requestId: msg.requestId, error: describeError(error) }, + }); + } + } + + private async handleDelete( + msg: Extract, + ): Promise { + try { + await this.requireRuntime().deleteLocalModel(parseModelId(msg.modelId)); + this.options.emit({ + type: "speech.local_models.delete.response", + payload: { requestId: msg.requestId, error: null }, + }); + } catch (error) { + this.options.emit({ + type: "speech.local_models.delete.response", + payload: { requestId: msg.requestId, error: describeError(error) }, + }); + } + } + + private requireRuntime(): SpeechModelsRuntime { + if (!this.options.runtime) { + throw new Error("Speech is unavailable on this daemon"); + } + return this.options.runtime; + } +} + +function parseModelId(modelId: string) { + const parsed = LocalSpeechModelIdSchema.safeParse(modelId); + if (!parsed.success) { + throw new Error(`Unknown local speech model: ${modelId}`); + } + return parsed.data; +} + +function toWireModel(model: LocalSpeechModelStatus): LocalSpeechModel { + const { spec, state, inUse } = model; + return { + id: spec.id, + kind: spec.kind === "tts" ? "tts" : "stt", + label: spec.label, + description: spec.description, + languages: [...spec.languages], + downloadSizeBytes: spec.downloadSizeBytes, + isDefault: spec.defaultFor !== undefined, + inUse, + status: state.status, + ...(state.status === "downloading" + ? { receivedBytes: state.receivedBytes, totalBytes: state.totalBytes } + : {}), + ...(state.status === "failed" ? { error: state.error } : {}), + }; +} + +function describeError(error: unknown): string { + return error instanceof Error ? error.message : String(error); +} diff --git a/packages/server/src/server/speech/providers/local/config.ts b/packages/server/src/server/speech/providers/local/config.ts index 8e7d500dc2..e15a8cae64 100644 --- a/packages/server/src/server/speech/providers/local/config.ts +++ b/packages/server/src/server/speech/providers/local/config.ts @@ -29,6 +29,7 @@ export interface LocalSpeechProviderConfig { export interface ResolvedLocalSpeechConfig { local: LocalSpeechProviderConfig | undefined; + modelsDir: string; sttLanguages: LocalSpeechSttLanguageConfig; } @@ -200,11 +201,10 @@ export function resolveLocalSpeechConfig(params: { buildLocalSpeechResolutionInput({ ...params, includeProviderConfig }), ); - const resolvedVoiceTtsSpeakerId = - parsed.voiceLocalTtsSpeakerId ?? - (parsed.voiceLocalTtsModel === "kokoro-en-v0_19" ? 0 : undefined); + const resolvedVoiceTtsSpeakerId = parsed.voiceLocalTtsSpeakerId ?? 0; return { + modelsDir: parsed.modelsDir, sttLanguages: { dictation: parsed.dictationLanguage, voice: parsed.voiceLanguage, diff --git a/packages/server/src/server/speech/providers/local/model-library.ts b/packages/server/src/server/speech/providers/local/model-library.ts new file mode 100644 index 0000000000..4ef19e3fa9 --- /dev/null +++ b/packages/server/src/server/speech/providers/local/model-library.ts @@ -0,0 +1,137 @@ +import type { Logger } from "pino"; + +import { + deleteSherpaOnnxModel, + ensureSherpaOnnxModel, + isSherpaOnnxModelInstalled, +} from "./sherpa/model-downloader.js"; +import { + listLocalSpeechModels, + type LocalSpeechModelId, + type LocalSpeechModelSpec, +} from "./models.js"; + +export type LocalSpeechModelInstallState = + | { status: "installed" } + | { status: "not_installed" } + | { status: "downloading"; receivedBytes: number; totalBytes: number } + | { status: "extracting" } + | { status: "failed"; error: string }; + +export interface LocalSpeechModelEntry { + spec: LocalSpeechModelSpec; + state: LocalSpeechModelInstallState; +} + +type InFlightState = Extract< + LocalSpeechModelInstallState, + { status: "downloading" | "extracting" } +>; + +interface InFlightDownload { + promise: Promise; + state: InFlightState; +} + +export class LocalSpeechModelBusyError extends Error { + constructor(public readonly modelId: LocalSpeechModelId) { + super(`${modelId} is downloading. Wait for the download to finish before deleting it.`); + this.name = "LocalSpeechModelBusyError"; + } +} + +// Owns the models directory: which catalog models are installed, which are downloading, +// and the last failure per model. Concurrent downloads of one model share a single transfer. +export class LocalSpeechModelLibrary { + private readonly inFlight = new Map(); + private readonly failures = new Map(); + private readonly abortController = new AbortController(); + private readonly logger: Logger; + + constructor( + private readonly modelsDir: string, + logger: Logger, + ) { + this.logger = logger.child({ module: "speech", component: "local-model-library" }); + } + + getModelsDir(): string { + return this.modelsDir; + } + + async list(): Promise { + return Promise.all( + listLocalSpeechModels().map(async (spec) => ({ spec, state: await this.getState(spec.id) })), + ); + } + + async listMissing(modelIds: LocalSpeechModelId[]): Promise { + const installed = await Promise.all( + modelIds.map((modelId) => isSherpaOnnxModelInstalled(this.modelsDir, modelId)), + ); + return modelIds.filter((_modelId, index) => !installed[index]); + } + + download(modelId: LocalSpeechModelId): Promise { + const existing = this.inFlight.get(modelId); + if (existing) { + return existing.promise; + } + this.failures.delete(modelId); + const entry: InFlightDownload = { + state: { status: "downloading", receivedBytes: 0, totalBytes: 0 }, + promise: Promise.resolve(""), + }; + entry.promise = ensureSherpaOnnxModel({ + modelsDir: this.modelsDir, + modelId, + logger: this.logger, + signal: this.abortController.signal, + onProgress: (progress) => { + entry.state = + progress.phase === "downloading" + ? { + status: "downloading", + receivedBytes: progress.receivedBytes, + totalBytes: progress.totalBytes, + } + : { status: "extracting" }; + }, + }) + .catch((error: unknown) => { + this.failures.set(modelId, error instanceof Error ? error.message : String(error)); + throw error; + }) + .finally(() => { + this.inFlight.delete(modelId); + }); + this.inFlight.set(modelId, entry); + return entry.promise; + } + + async delete(modelId: LocalSpeechModelId): Promise { + if (this.inFlight.has(modelId)) { + throw new LocalSpeechModelBusyError(modelId); + } + this.failures.delete(modelId); + await deleteSherpaOnnxModel(this.modelsDir, modelId); + this.logger.info({ modelId }, "Deleted local speech model"); + } + + async stop(): Promise { + this.abortController.abort(); + await Promise.allSettled(Array.from(this.inFlight.values(), (entry) => entry.promise)); + } + + private async getState(modelId: LocalSpeechModelId): Promise { + const inFlight = this.inFlight.get(modelId); + if (inFlight) { + return inFlight.state; + } + if (await isSherpaOnnxModelInstalled(this.modelsDir, modelId)) { + return { status: "installed" }; + } + const failure = this.failures.get(modelId); + return failure ? { status: "failed", error: failure } : { status: "not_installed" }; + } +} diff --git a/packages/server/src/server/speech/providers/local/models.test.ts b/packages/server/src/server/speech/providers/local/models.test.ts index e40cc25384..be010dcae9 100644 --- a/packages/server/src/server/speech/providers/local/models.test.ts +++ b/packages/server/src/server/speech/providers/local/models.test.ts @@ -13,12 +13,8 @@ import path from "node:path"; import pino from "pino"; import { describe, expect, onTestFinished, test } from "vitest"; -import { - ensureLocalSpeechModels, - getLocalSpeechModelDir, - listLocalSpeechModels, - listMissingLocalSpeechModels, -} from "./models.js"; +import { LocalSpeechModelLibrary } from "./model-library.js"; +import { getLocalSpeechModelDir, listLocalSpeechModels } from "./models.js"; const logger = pino({ level: "silent" }); const modelId = "parakeet-tdt-0.6b-v2-int8"; @@ -55,16 +51,31 @@ describe("local speech models", () => { const encoder = path.join(modelDir, spec.requiredFiles[0]); truncateSync(encoder, 1_000); - expect(await listMissingLocalSpeechModels({ modelsDir, modelIds: [modelId] })).toEqual([ - modelId, - ]); + const library = new LocalSpeechModelLibrary(modelsDir, logger); + expect(await library.listMissing([modelId])).toEqual([modelId]); - await ensureLocalSpeechModels({ modelsDir, modelIds: [modelId], logger }); + await library.download(modelId); expect(readFileSync(encoder, "utf8")).toBe( `complete ${spec.requiredFiles[0]}`.padEnd(64_000, "."), ); expect(existsSync(archive)).toBe(false); - expect(await listMissingLocalSpeechModels({ modelsDir, modelIds: [modelId] })).toEqual([]); + expect(await library.listMissing([modelId])).toEqual([]); + }); + + test("lists install state and deletes an installed model", async () => { + const { modelsDir } = setupModelsDirWithArchive(); + const library = new LocalSpeechModelLibrary(modelsDir, logger); + const stateOf = async () => + (await library.list()).find((entry) => entry.spec.id === modelId)?.state; + + expect(await stateOf()).toEqual({ status: "not_installed" }); + await library.download(modelId); + expect(await stateOf()).toEqual({ status: "installed" }); + + await library.delete(modelId); + + expect(await stateOf()).toEqual({ status: "not_installed" }); + expect(existsSync(getLocalSpeechModelDir(modelsDir, modelId))).toBe(false); }); }); diff --git a/packages/server/src/server/speech/providers/local/models.ts b/packages/server/src/server/speech/providers/local/models.ts index a884be86d7..7d4c7f56b5 100644 --- a/packages/server/src/server/speech/providers/local/models.ts +++ b/packages/server/src/server/speech/providers/local/models.ts @@ -1,11 +1,8 @@ -import { - ensureSherpaOnnxModels, - getSherpaOnnxModelDir, - isSherpaOnnxModelInstalled, -} from "./sherpa/model-downloader.js"; +import { ensureSherpaOnnxModels, getSherpaOnnxModelDir } from "./sherpa/model-downloader.js"; import { DEFAULT_LOCAL_STT_MODEL, DEFAULT_LOCAL_TTS_MODEL, + LocalSpeechModelIdSchema, LocalSttModelIdSchema, LocalTtsModelIdSchema, listSherpaOnnxModels, @@ -17,6 +14,7 @@ import { export { DEFAULT_LOCAL_STT_MODEL, DEFAULT_LOCAL_TTS_MODEL, + LocalSpeechModelIdSchema, LocalSttModelIdSchema, LocalTtsModelIdSchema, type LocalSpeechModelId, @@ -34,16 +32,6 @@ export function getLocalSpeechModelDir(modelsDir: string, modelId: LocalSpeechMo return getSherpaOnnxModelDir(modelsDir, modelId); } -export async function listMissingLocalSpeechModels(options: { - modelsDir: string; - modelIds: LocalSpeechModelId[]; -}): Promise { - const installed = await Promise.all( - options.modelIds.map((modelId) => isSherpaOnnxModelInstalled(options.modelsDir, modelId)), - ); - return options.modelIds.filter((_modelId, index) => !installed[index]); -} - export async function ensureLocalSpeechModels(options: { modelsDir: string; modelIds: LocalSpeechModelId[]; diff --git a/packages/server/src/server/speech/providers/local/sherpa/model-catalog.ts b/packages/server/src/server/speech/providers/local/sherpa/model-catalog.ts index 06d838c49a..b142f52dd6 100644 --- a/packages/server/src/server/speech/providers/local/sherpa/model-catalog.ts +++ b/packages/server/src/server/speech/providers/local/sherpa/model-catalog.ts @@ -4,43 +4,467 @@ export type SherpaOnnxModelKind = "stt-offline" | "tts"; type DefaultModelRole = "stt" | "tts"; -interface SherpaOnnxCatalogEntry { - kind: SherpaOnnxModelKind; +// How sherpa-onnx loads a model, with file names relative to the extracted directory. +export type SttEngine = + | { family: "transducer"; encoder: string; decoder: string; joiner: string; tokens: string } + | { family: "whisper"; encoder: string; decoder: string; tokens: string } + | { family: "sense_voice"; model: string; tokens: string } + | { family: "dolphin"; model: string; tokens: string } + | { family: "nemo_ctc"; model: string; tokens: string }; + +export type TtsEngine = + | { + family: "kokoro"; + model: string; + voices: string; + tokens: string; + dataDir: string; + // Multi-language Kokoro models read pronunciations from these instead of espeak alone. + lexicons?: readonly string[]; + dictDir?: string; + ruleFsts?: readonly string[]; + } + | { family: "kitten"; model: string; voices: string; tokens: string; dataDir: string } + | { family: "vits"; model: string; tokens: string; dataDir: string }; + +interface SherpaOnnxCatalogEntryBase { + label: string; + // One user-facing sentence shown before download. + description: string; + // BCP 47 language tags the model transcribes or speaks. + languages: readonly string[]; + // Exact size of the release archive. The downloader rejects any other size. + downloadSizeBytes: number; archiveUrl: string; extractedDir: string; - requiredFiles: string[]; - description: string; defaultFor?: DefaultModelRole; } +type SherpaOnnxCatalogEntry = + | (SherpaOnnxCatalogEntryBase & { kind: "stt-offline"; engine: SttEngine }) + | (SherpaOnnxCatalogEntryBase & { kind: "tts"; engine: TtsEngine }); + +// Every sherpa-onnx release archive extracts to a directory named after the archive. +function releaseArchive(tag: "asr-models" | "tts-models", name: string) { + return { + archiveUrl: `https://github.com/k2-fsa/sherpa-onnx/releases/download/${tag}/${name}.tar.bz2`, + extractedDir: name, + }; +} + +const PARAKEET_INT8_ENGINE = { + family: "transducer", + encoder: "encoder.int8.onnx", + decoder: "decoder.int8.onnx", + joiner: "joiner.int8.onnx", + tokens: "tokens.txt", +} as const satisfies SttEngine; + +function whisperEngine(size: string): SttEngine { + return { + family: "whisper", + encoder: `${size}-encoder.int8.onnx`, + decoder: `${size}-decoder.int8.onnx`, + tokens: `${size}-tokens.txt`, + }; +} + +function piperEngine(voice: string): TtsEngine { + return { + family: "vits", + model: `${voice}.onnx`, + tokens: "tokens.txt", + dataDir: "espeak-ng-data", + }; +} + +// Parakeet TDT v3 transcribes these 25 European languages and detects which one is spoken. +const PARAKEET_V3_LANGUAGES = [ + "bg", + "hr", + "cs", + "da", + "nl", + "en", + "et", + "fi", + "fr", + "de", + "el", + "hu", + "it", + "lv", + "lt", + "mt", + "pl", + "pt", + "ro", + "sk", + "sl", + "es", + "sv", + "ru", + "uk", +] as const; + +// The 99 languages multilingual Whisper models transcribe, detecting which one is spoken. +const WHISPER_LANGUAGES = [ + "en", + "zh", + "de", + "es", + "ru", + "ko", + "fr", + "ja", + "pt", + "tr", + "pl", + "ca", + "nl", + "ar", + "sv", + "it", + "id", + "hi", + "fi", + "vi", + "he", + "uk", + "el", + "ms", + "cs", + "ro", + "da", + "hu", + "ta", + "no", + "th", + "ur", + "hr", + "bg", + "lt", + "la", + "mi", + "ml", + "cy", + "sk", + "te", + "fa", + "lv", + "bn", + "sr", + "az", + "sl", + "kn", + "et", + "mk", + "br", + "eu", + "is", + "hy", + "ne", + "mn", + "bs", + "kk", + "sq", + "sw", + "gl", + "mr", + "pa", + "si", + "km", + "sn", + "yo", + "so", + "af", + "oc", + "ka", + "be", + "tg", + "sd", + "gu", + "am", + "yi", + "lo", + "uz", + "fo", + "ht", + "ps", + "tk", + "nn", + "mt", + "sa", + "lb", + "my", + "bo", + "tl", + "mg", + "as", + "tt", + "haw", + "ln", + "ha", + "ba", + "jv", + "su", +] as const; + +// Dolphin's 40 Eastern languages. It also covers 22 Chinese dialects under "zh". +const DOLPHIN_LANGUAGES = [ + "zh", + "ja", + "th", + "ru", + "ko", + "id", + "vi", + "yue", + "hi", + "ur", + "ms", + "uz", + "ar", + "fa", + "bn", + "ta", + "te", + "ug", + "gu", + "my", + "tl", + "kk", + "or", + "ne", + "mn", + "km", + "jv", + "lo", + "si", + "fil", + "ps", + "pa", + "kab", + "ba", + "ks", + "tg", + "su", + "mr", + "ky", + "az", +] as const; + export const SHERPA_ONNX_MODEL_CATALOG = { "parakeet-tdt-0.6b-v2-int8": { kind: "stt-offline", - archiveUrl: - "https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8.tar.bz2", - extractedDir: "sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8", - requiredFiles: ["encoder.int8.onnx", "decoder.int8.onnx", "joiner.int8.onnx", "tokens.txt"], - description: "NVIDIA Parakeet TDT v2 (offline NeMo transducer, English).", + label: "Parakeet v2", + description: "Accurate English transcription that runs well on a laptop CPU.", + languages: ["en"], + downloadSizeBytes: 482_468_385, + ...releaseArchive("asr-models", "sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8"), + engine: PARAKEET_INT8_ENGINE, defaultFor: "stt", }, + "parakeet-tdt-110m-en-int8": { + kind: "stt-offline", + label: "Parakeet Small", + description: "Small, fast English model with punctuation. A quick download for any laptop.", + languages: ["en"], + downloadSizeBytes: 108_035_095, + ...releaseArchive("asr-models", "sherpa-onnx-nemo-parakeet_tdt_transducer_110m-en-36000-int8"), + engine: PARAKEET_INT8_ENGINE, + }, "parakeet-tdt-0.6b-v3-int8": { kind: "stt-offline", - archiveUrl: - "https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8.tar.bz2", - extractedDir: "sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8", - requiredFiles: ["encoder.int8.onnx", "decoder.int8.onnx", "joiner.int8.onnx", "tokens.txt"], - description: - "NVIDIA Parakeet TDT v3 (offline NeMo transducer, 25 European languages, auto-detected).", + label: "Parakeet v3", + description: "Parakeet for 25 European languages, detecting which one you speak.", + languages: PARAKEET_V3_LANGUAGES, + downloadSizeBytes: 487_170_055, + ...releaseArchive("asr-models", "sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8"), + engine: PARAKEET_INT8_ENGINE, + }, + "whisper-tiny-en": { + kind: "stt-offline", + label: "Whisper Tiny (English)", + description: "Small English Whisper model. Quick to download, less accurate than Parakeet.", + languages: ["en"], + downloadSizeBytes: 118_071_777, + ...releaseArchive("asr-models", "sherpa-onnx-whisper-tiny.en"), + engine: whisperEngine("tiny.en"), + }, + "whisper-base": { + kind: "stt-offline", + label: "Whisper Base", + description: "Compact Whisper model for 99 languages, detecting which one you speak.", + languages: WHISPER_LANGUAGES, + downloadSizeBytes: 207_557_382, + ...releaseArchive("asr-models", "sherpa-onnx-whisper-base"), + engine: whisperEngine("base"), + }, + "whisper-turbo": { + kind: "stt-offline", + label: "Whisper Turbo", + description: "Whisper's most accurate model for 99 languages and Cantonese. Slower on a CPU.", + languages: [...WHISPER_LANGUAGES, "yue"], + downloadSizeBytes: 563_790_207, + ...releaseArchive("asr-models", "sherpa-onnx-whisper-turbo"), + engine: whisperEngine("turbo"), + }, + "sense-voice-int8": { + kind: "stt-offline", + label: "SenseVoice", + description: "Chinese, Cantonese, English, Japanese and Korean, detecting which one you speak.", + languages: ["zh", "yue", "en", "ja", "ko"], + downloadSizeBytes: 165_783_878, + ...releaseArchive("asr-models", "sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2025-09-09"), + engine: { family: "sense_voice", model: "model.int8.onnx", tokens: "tokens.txt" }, + }, + "dolphin-base-int8": { + kind: "stt-offline", + label: "Dolphin", + description: "Small model for 40 Asian and Middle Eastern languages and Chinese dialects.", + languages: DOLPHIN_LANGUAGES, + downloadSizeBytes: 80_671_385, + ...releaseArchive("asr-models", "sherpa-onnx-dolphin-base-ctc-multi-lang-int8-2025-04-02"), + engine: { family: "dolphin", model: "model.int8.onnx", tokens: "tokens.txt" }, + }, + "giga-am-v3-ru": { + kind: "stt-offline", + label: "GigaAM v3 (Russian)", + description: "Russian transcription with punctuation.", + languages: ["ru"], + downloadSizeBytes: 170_197_019, + ...releaseArchive( + "asr-models", + "sherpa-onnx-nemo-transducer-punct-giga-am-v3-russian-2025-12-16", + ), + engine: { + family: "transducer", + encoder: "encoder.int8.onnx", + decoder: "decoder.onnx", + joiner: "joiner.onnx", + tokens: "tokens.txt", + }, + }, + "parakeet-tdt-ctc-0.6b-ja-int8": { + kind: "stt-offline", + label: "Parakeet (Japanese)", + description: "Japanese transcription with punctuation.", + languages: ["ja"], + downloadSizeBytes: 489_389_564, + ...releaseArchive("asr-models", "sherpa-onnx-nemo-parakeet-tdt_ctc-0.6b-ja-35000-int8"), + engine: { family: "nemo_ctc", model: "model.int8.onnx", tokens: "tokens.txt" }, }, "kokoro-en-v0_19": { kind: "tts", - archiveUrl: - "https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/kokoro-en-v0_19.tar.bz2", - extractedDir: "kokoro-en-v0_19", - requiredFiles: ["model.onnx", "voices.bin", "tokens.txt", "espeak-ng-data"], - description: "Kokoro TTS (higher quality; larger).", + label: "Kokoro (English)", + description: "Natural-sounding English voice for voice mode replies.", + languages: ["en"], + downloadSizeBytes: 319_625_534, + ...releaseArchive("tts-models", "kokoro-en-v0_19"), + engine: { + family: "kokoro", + model: "model.onnx", + voices: "voices.bin", + tokens: "tokens.txt", + dataDir: "espeak-ng-data", + }, defaultFor: "tts", }, + "kokoro-int8-en-v0_19": { + kind: "tts", + label: "Kokoro Lite (English)", + description: "The same Kokoro English voices, quantized to a third of the download.", + languages: ["en"], + downloadSizeBytes: 103_248_205, + ...releaseArchive("tts-models", "kokoro-int8-en-v0_19"), + engine: { + family: "kokoro", + model: "model.int8.onnx", + voices: "voices.bin", + tokens: "tokens.txt", + dataDir: "espeak-ng-data", + }, + }, + "kokoro-int8-multi-lang-v1_1": { + kind: "tts", + label: "Kokoro (Chinese and English)", + description: "Kokoro voices that speak both Mandarin Chinese and English.", + languages: ["zh", "en"], + downloadSizeBytes: 147_031_220, + ...releaseArchive("tts-models", "kokoro-int8-multi-lang-v1_1"), + engine: { + family: "kokoro", + model: "model.int8.onnx", + voices: "voices.bin", + tokens: "tokens.txt", + dataDir: "espeak-ng-data", + lexicons: ["lexicon-us-en.txt", "lexicon-zh.txt"], + dictDir: "dict", + ruleFsts: ["date-zh.fst", "phone-zh.fst", "number-zh.fst"], + }, + }, + // Kitten v0.8 needs a newer sherpa-onnx than we pin; v0.2 loads on 1.12.28. + "kitten-nano-en-v0_2-fp16": { + kind: "tts", + label: "Kitten Nano (English)", + description: "Tiny, very fast English voice.", + languages: ["en"], + downloadSizeBytes: 26_586_708, + ...releaseArchive("tts-models", "kitten-nano-en-v0_2-fp16"), + engine: { + family: "kitten", + model: "model.fp16.onnx", + voices: "voices.bin", + tokens: "tokens.txt", + dataDir: "espeak-ng-data", + }, + }, + "piper-en-us-libritts-r": { + kind: "tts", + label: "Piper LibriTTS (US English)", + description: "Lightweight American English voice.", + languages: ["en-US"], + downloadSizeBytes: 23_398_348, + ...releaseArchive("tts-models", "vits-piper-en_US-libritts_r-medium-int8"), + engine: piperEngine("en_US-libritts_r-medium"), + }, + "piper-en-gb-alba": { + kind: "tts", + label: "Piper Alba (British English)", + description: "Lightweight British English voice.", + languages: ["en-GB"], + downloadSizeBytes: 21_104_326, + ...releaseArchive("tts-models", "vits-piper-en_GB-alba-medium-int8"), + engine: piperEngine("en_GB-alba-medium"), + }, + "piper-de-thorsten": { + kind: "tts", + label: "Piper Thorsten (German)", + description: "Lightweight German voice.", + languages: ["de"], + downloadSizeBytes: 20_949_833, + ...releaseArchive("tts-models", "vits-piper-de_DE-thorsten-medium-int8"), + engine: piperEngine("de_DE-thorsten-medium"), + }, + "piper-fr-siwis": { + kind: "tts", + label: "Piper Siwis (French)", + description: "Lightweight French voice.", + languages: ["fr"], + downloadSizeBytes: 20_914_888, + ...releaseArchive("tts-models", "vits-piper-fr_FR-siwis-medium-int8"), + engine: piperEngine("fr_FR-siwis-medium"), + }, + "piper-es-davefx": { + kind: "tts", + label: "Piper Davefx (Spanish)", + description: "Lightweight Spanish voice.", + languages: ["es"], + downloadSizeBytes: 21_171_632, + ...releaseArchive("tts-models", "vits-piper-es_ES-davefx-medium-int8"), + engine: piperEngine("es_ES-davefx-medium"), + }, } as const satisfies Record; export type SherpaOnnxModelId = keyof typeof SHERPA_ONNX_MODEL_CATALOG; @@ -97,24 +521,44 @@ function createModelIdSchema(modelIds: readonly T[]): z.ZodTyp .transform((value) => value as T); } +export const LocalSpeechModelIdSchema = createModelIdSchema(ALL_MODEL_IDS); export const LocalSttModelIdSchema = createModelIdSchema(LOCAL_STT_MODEL_IDS); export const LocalTtsModelIdSchema = createModelIdSchema(LOCAL_TTS_MODEL_IDS); export type SherpaOnnxModelSpec = SherpaOnnxCatalogEntry & { id: SherpaOnnxModelId; + // Every file and directory that must exist for the model to count as installed. + requiredFiles: string[]; }; export function listSherpaOnnxModels(): SherpaOnnxModelSpec[] { - return ALL_MODEL_IDS.map((id) => Object.assign({ id }, SHERPA_ONNX_MODEL_CATALOG[id])); + return ALL_MODEL_IDS.map(getSherpaOnnxModelSpec); } export function getSherpaOnnxModelSpec(id: SherpaOnnxModelId): SherpaOnnxModelSpec { - const spec = SHERPA_ONNX_MODEL_CATALOG[id]; + const spec: SherpaOnnxCatalogEntry | undefined = SHERPA_ONNX_MODEL_CATALOG[id]; if (!spec) { throw new Error(`Unknown local speech model id: ${id}`); } return { id, ...spec, + requiredFiles: engineFiles(spec.engine), }; } + +// Every path an engine names, so the downloader can verify an install. +export function engineFiles(engine: SttEngine | TtsEngine): string[] { + return Object.entries(engine).flatMap(([key, value]) => { + if (key === "family") return []; + return typeof value === "string" ? [value] : (value as readonly string[]); + }); +} + +export function getLocalSttEngine(id: LocalSttModelId): SttEngine { + return SHERPA_ONNX_MODEL_CATALOG[id].engine; +} + +export function getLocalTtsEngine(id: LocalTtsModelId): TtsEngine { + return SHERPA_ONNX_MODEL_CATALOG[id].engine; +} diff --git a/packages/server/src/server/speech/providers/local/sherpa/model-downloader.ts b/packages/server/src/server/speech/providers/local/sherpa/model-downloader.ts index cdd286389f..b7ba511224 100644 --- a/packages/server/src/server/speech/providers/local/sherpa/model-downloader.ts +++ b/packages/server/src/server/speech/providers/local/sherpa/model-downloader.ts @@ -8,11 +8,16 @@ import type pino from "pino"; import { getSherpaOnnxModelSpec, type SherpaOnnxModelId } from "./model-catalog.js"; import { spawnProcess } from "../../../../../utils/spawn.js"; +export type SherpaOnnxModelInstallProgress = + | { phase: "downloading"; receivedBytes: number; totalBytes: number } + | { phase: "extracting" }; + export interface EnsureSherpaOnnxModelOptions { modelsDir: string; modelId: SherpaOnnxModelId; logger: pino.Logger; signal?: AbortSignal; + onProgress?: (progress: SherpaOnnxModelInstallProgress) => void; } export function getSherpaOnnxModelDir(modelsDir: string, modelId: SherpaOnnxModelId): string { @@ -41,7 +46,9 @@ async function hasRequiredFiles(modelDir: string, requiredFiles: string[]): Prom interface DownloadToFileOptions { url: string; outputPath: string; + expectedBytes: number; signal?: AbortSignal; + onProgress?: (receivedBytes: number, totalBytes: number) => void; } async function downloadToFile(options: DownloadToFileOptions): Promise { @@ -60,6 +67,13 @@ async function downloadToFile(options: DownloadToFileOptions): Promise { // The fetch ReadableStream type is slightly different from what Readable.fromWeb expects // eslint-disable-next-line @typescript-eslint/no-explicit-any const nodeStream = Readable.fromWeb(res.body as any); + const contentLength = Number(res.headers.get("content-length")); + const totalBytes = contentLength > 0 ? contentLength : options.expectedBytes; + let receivedBytes = 0; + nodeStream.on("data", (chunk: Buffer) => { + receivedBytes += chunk.length; + options.onProgress?.(receivedBytes, totalBytes); + }); try { await pipeline(nodeStream, createWriteStream(tmpPath), { signal: options.signal }); @@ -157,9 +171,13 @@ export async function ensureSherpaOnnxModel( await downloadToFile({ url: spec.archiveUrl, outputPath: archivePath, + expectedBytes: spec.downloadSizeBytes, signal: options.signal, + onProgress: (receivedBytes, totalBytes) => + options.onProgress?.({ phase: "downloading", receivedBytes, totalBytes }), }); } + options.onProgress?.({ phase: "extracting" }); logger.info( { @@ -213,6 +231,16 @@ export async function ensureSherpaOnnxModel( } } +export async function deleteSherpaOnnxModel( + modelsDir: string, + modelId: SherpaOnnxModelId, +): Promise { + const { modelDir, archivePath, extractionDir } = getSherpaOnnxModelPaths(modelsDir, modelId); + await rm(modelDir, { recursive: true, force: true }); + await rm(archivePath, { force: true }); + await rm(extractionDir, { recursive: true, force: true }); +} + export async function ensureSherpaOnnxModels(options: { modelsDir: string; modelIds: SherpaOnnxModelId[]; diff --git a/packages/server/src/server/speech/providers/local/sherpa/sherpa-offline-recognizer.ts b/packages/server/src/server/speech/providers/local/sherpa/sherpa-offline-recognizer.ts index cfde73addf..5137f51810 100644 --- a/packages/server/src/server/speech/providers/local/sherpa/sherpa-offline-recognizer.ts +++ b/packages/server/src/server/speech/providers/local/sherpa/sherpa-offline-recognizer.ts @@ -1,6 +1,8 @@ import { existsSync } from "node:fs"; +import path from "node:path"; import type pino from "pino"; +import { engineFiles, type SttEngine } from "./model-catalog.js"; import { loadSherpaOnnxNode } from "./sherpa-onnx-node-loader.js"; function assertFileExists(filePath: string, label: string): void { @@ -9,16 +11,43 @@ function assertFileExists(filePath: string, label: string): void { } } -export interface SherpaOfflineRecognizerModel { - kind: "nemo_transducer"; - encoder: string; - decoder: string; - joiner: string; - tokens: string; +// Each engine family takes its own block under sherpa's modelConfig. +function buildModelConfig(modelDir: string, engine: SttEngine): Record { + const file = (name: string) => path.join(modelDir, name); + switch (engine.family) { + case "transducer": + return { + transducer: { + encoder: file(engine.encoder), + decoder: file(engine.decoder), + joiner: file(engine.joiner), + }, + modelType: "nemo_transducer", + }; + case "whisper": + // An empty language makes Whisper detect the spoken language. + return { + whisper: { + encoder: file(engine.encoder), + decoder: file(engine.decoder), + language: "", + task: "transcribe", + }, + }; + case "sense_voice": + return { + senseVoice: { model: file(engine.model), language: "", useInverseTextNormalization: 1 }, + }; + case "dolphin": + return { dolphin: { model: file(engine.model) } }; + case "nemo_ctc": + return { nemoCtc: { model: file(engine.model) } }; + } } export interface SherpaOfflineRecognizerConfig { - model: SherpaOfflineRecognizerModel; + modelDir: string; + engine: SttEngine; numThreads?: number; provider?: "cpu"; debug?: 0 | 1; @@ -54,10 +83,9 @@ export class SherpaOfflineRecognizerEngine { component: "offline-recognizer", }); - assertFileExists(config.model.encoder, "offline encoder"); - assertFileExists(config.model.decoder, "offline decoder"); - assertFileExists(config.model.joiner, "offline joiner"); - assertFileExists(config.model.tokens, "tokens"); + for (const name of engineFiles(config.engine)) { + assertFileExists(path.join(config.modelDir, name), `${config.engine.family} model file`); + } const sherpa = loadSherpaOnnxNode(); @@ -67,13 +95,8 @@ export class SherpaOfflineRecognizerEngine { featureDim: config.featureDim ?? 80, }, modelConfig: { - transducer: { - encoder: config.model.encoder, - decoder: config.model.decoder, - joiner: config.model.joiner, - }, - tokens: config.model.tokens, - modelType: "nemo_transducer", + ...buildModelConfig(config.modelDir, config.engine), + tokens: path.join(config.modelDir, config.engine.tokens), numThreads: config.numThreads ?? 1, provider: config.provider ?? "cpu", debug: config.debug ?? 0, diff --git a/packages/server/src/server/speech/providers/local/sherpa/sherpa-tts.test.ts b/packages/server/src/server/speech/providers/local/sherpa/sherpa-tts.test.ts index 06d4a5b41d..0ae51f86ee 100644 --- a/packages/server/src/server/speech/providers/local/sherpa/sherpa-tts.test.ts +++ b/packages/server/src/server/speech/providers/local/sherpa/sherpa-tts.test.ts @@ -3,6 +3,7 @@ import { mkdtempSync, mkdirSync, rmSync, writeFileSync } from "node:fs"; import { tmpdir } from "node:os"; import { join } from "node:path"; import pino from "pino"; +import { engineFiles, getLocalTtsEngine, type LocalTtsModelId } from "./model-catalog.js"; import { SherpaOnnxTTS } from "./sherpa-tts.js"; const directories: string[] = []; @@ -11,17 +12,28 @@ afterEach(() => { rmSync(directory, { recursive: true, force: true }); }); -function createTts(numThreads?: number, beforeGenerate?: () => Promise) { +function createTts( + numThreads?: number, + beforeGenerate?: () => Promise, + modelId: LocalTtsModelId = "kokoro-en-v0_19", +) { const modelDir = mkdtempSync(join(tmpdir(), "paseo-tts-")); directories.push(modelDir); - for (const name of ["model.onnx", "voices.bin", "tokens.txt"]) - writeFileSync(join(modelDir, name), ""); - mkdirSync(join(modelDir, "espeak-ng-data")); + // Data directories have no extension; everything else is a model file. + for (const name of engineFiles(getLocalTtsEngine(modelId))) { + if (name.includes(".")) writeFileSync(join(modelDir, name), ""); + else mkdirSync(join(modelDir, name)); + } let nativeConfig: unknown; const requests: unknown[] = []; const released: boolean[] = []; const tts = new SherpaOnnxTTS( - { preset: "kokoro-en-v0_19", modelDir, numThreads }, + { + modelId, + modelDir, + engine: getLocalTtsEngine(modelId), + numThreads, + }, pino({ level: "silent" }), () => ({ OfflineTts: class { @@ -138,4 +150,31 @@ describe("SherpaOnnxTTS", () => { expect(nativeConfig).not.toHaveProperty("provider"); }, ); + + it("loads Piper voices through the vits engine", () => { + const { nativeConfig } = createTts(undefined, undefined, "piper-de-thorsten"); + expect(nativeConfig).toMatchObject({ + model: { + vits: { + model: expect.stringMatching(/de_DE-thorsten-medium\.onnx$/), + tokens: expect.stringMatching(/tokens\.txt$/), + dataDir: expect.stringMatching(/espeak-ng-data$/), + }, + }, + }); + expect(nativeConfig).not.toHaveProperty("model.kokoro"); + }); + + it("gives multi-language Kokoro its lexicons, dictionary and number rules", () => { + const { nativeConfig } = createTts(undefined, undefined, "kokoro-int8-multi-lang-v1_1"); + expect(nativeConfig).toMatchObject({ + model: { + kokoro: { + lexicon: expect.stringMatching(/lexicon-us-en\.txt,.*lexicon-zh\.txt$/), + dictDir: expect.stringMatching(/dict$/), + }, + }, + ruleFsts: expect.stringMatching(/date-zh\.fst,.*phone-zh\.fst,.*number-zh\.fst$/), + }); + }); }); diff --git a/packages/server/src/server/speech/providers/local/sherpa/sherpa-tts.ts b/packages/server/src/server/speech/providers/local/sherpa/sherpa-tts.ts index a078f81f51..12c86b8777 100644 --- a/packages/server/src/server/speech/providers/local/sherpa/sherpa-tts.ts +++ b/packages/server/src/server/speech/providers/local/sherpa/sherpa-tts.ts @@ -1,16 +1,17 @@ import type pino from "pino"; import { Readable } from "node:stream"; import { existsSync } from "node:fs"; +import path from "node:path"; import type { SpeechStreamResult, TextToSpeechProvider } from "../../../speech-provider.js"; import { chunkBuffer, float32ToPcm16le } from "../../../audio.js"; import { loadSherpaOnnxNode } from "./sherpa-onnx-node-loader.js"; - -export type SherpaTtsPreset = "kokoro-en-v0_19"; +import { engineFiles, type LocalTtsModelId, type TtsEngine } from "./model-catalog.js"; export interface SherpaTtsConfig { - preset: SherpaTtsPreset; + modelId: LocalTtsModelId; modelDir: string; + engine: TtsEngine; speakerId?: number; speed?: number; lengthScale?: number; @@ -23,6 +24,53 @@ function assertFileExists(filePath: string, label: string): void { } } +// Each engine family takes its own block under sherpa's TTS model config. +function buildTtsModelConfig( + modelDir: string, + engine: TtsEngine, + lengthScale: number, +): Record { + const file = (name: string) => path.join(modelDir, name); + switch (engine.family) { + case "kokoro": + return { + kokoro: { + model: file(engine.model), + voices: file(engine.voices), + tokens: file(engine.tokens), + dataDir: file(engine.dataDir), + lexicon: (engine.lexicons ?? []).map(file).join(","), + dictDir: engine.dictDir ? file(engine.dictDir) : "", + lengthScale, + }, + }; + case "kitten": + return { + kitten: { + model: file(engine.model), + voices: file(engine.voices), + tokens: file(engine.tokens), + dataDir: file(engine.dataDir), + lengthScale, + }, + }; + case "vits": + return { + vits: { + model: file(engine.model), + tokens: file(engine.tokens), + dataDir: file(engine.dataDir), + lengthScale, + }, + }; + } +} + +function ruleFsts(modelDir: string, engine: TtsEngine): string { + if (engine.family !== "kokoro" || !engine.ruleFsts) return ""; + return engine.ruleFsts.map((name) => path.join(modelDir, name)).join(","); +} + interface SherpaOfflineTtsNative { sampleRate?: number; generateAsync: (args: { @@ -59,39 +107,28 @@ export class SherpaOnnxTTS implements TextToSpeechProvider { throw new Error("sherpa-onnx-node OfflineTts is unavailable"); } - const modelPath = `${config.modelDir}/model.onnx`; - const voicesPath = `${config.modelDir}/voices.bin`; - const tokensPath = `${config.modelDir}/tokens.txt`; - const dataDir = `${config.modelDir}/espeak-ng-data`; - - assertFileExists(modelPath, "TTS model"); - assertFileExists(voicesPath, "TTS voices"); - assertFileExists(tokensPath, "TTS tokens"); - assertFileExists(dataDir, "TTS espeak-ng dataDir"); + for (const name of engineFiles(config.engine)) { + assertFileExists(path.join(config.modelDir, name), `${config.engine.family} TTS file`); + } const modelConfig = { // The native parser reads these under model, despite the upstream JS typedef. numThreads: config.numThreads ?? 2, provider: "cpu", - kokoro: { - model: modelPath, - voices: voicesPath, - tokens: tokensPath, - dataDir, - lengthScale: config.lengthScale ?? 1.0, - }, + ...buildTtsModelConfig(config.modelDir, config.engine, config.lengthScale ?? 1.0), }; const offlineTtsConfig = { model: modelConfig, maxNumSentences: 1, + ruleFsts: ruleFsts(config.modelDir, config.engine), }; this.tts = new ( sherpa as unknown as { OfflineTts: new (config: unknown) => SherpaOfflineTtsNative } ).OfflineTts(offlineTtsConfig); this.logger.info( - { preset: config.preset, modelDir: config.modelDir }, + { modelId: config.modelId, modelDir: config.modelDir }, "Sherpa offline TTS initialized", ); } diff --git a/packages/server/src/server/speech/providers/local/sherpa/speech-download.e2e.test.ts b/packages/server/src/server/speech/providers/local/sherpa/speech-download.e2e.test.ts index ab2eefef55..43ab8af01a 100644 --- a/packages/server/src/server/speech/providers/local/sherpa/speech-download.e2e.test.ts +++ b/packages/server/src/server/speech/providers/local/sherpa/speech-download.e2e.test.ts @@ -7,6 +7,7 @@ import pino from "pino"; import { ensureSherpaOnnxModels, getSherpaOnnxModelDir } from "./model-downloader.js"; import { createDaemonTestContext } from "../../../../test-utils/index.js"; import { parsePcm16MonoWav, wordSimilarity } from "../../../../test-utils/dictation-e2e.js"; +import { getLocalSttEngine, getLocalTtsEngine } from "./model-catalog.js"; import { SherpaOnnxTTS } from "./sherpa-tts.js"; import { SherpaOfflineRecognizerEngine } from "./sherpa-offline-recognizer.js"; import { SherpaOnnxParakeetSTT } from "./sherpa-parakeet-stt.js"; @@ -192,7 +193,12 @@ downloadTest( const ttsText = "This is a voice note."; const ttsModelDir = getSherpaOnnxModelDir(modelsDir, "kokoro-en-v0_19"); const tts = new SherpaOnnxTTS( - { preset: "kokoro-en-v0_19", modelDir: ttsModelDir, speakerId: 0 }, + { + modelId: "kokoro-en-v0_19", + modelDir: ttsModelDir, + engine: getLocalTtsEngine("kokoro-en-v0_19"), + speakerId: 0, + }, logger, ); const { stream, format: ttsFormat } = await tts.synthesizeSpeech(ttsText); @@ -206,13 +212,8 @@ downloadTest( const sttModelDir = getSherpaOnnxModelDir(modelsDir, "parakeet-tdt-0.6b-v2-int8"); const engine = new SherpaOfflineRecognizerEngine( { - model: { - kind: "nemo_transducer", - encoder: `${sttModelDir}/encoder.int8.onnx`, - decoder: `${sttModelDir}/decoder.int8.onnx`, - joiner: `${sttModelDir}/joiner.int8.onnx`, - tokens: `${sttModelDir}/tokens.txt`, - }, + modelDir: sttModelDir, + engine: getLocalSttEngine("parakeet-tdt-0.6b-v2-int8"), numThreads: 2, debug: 0, }, diff --git a/packages/server/src/server/speech/providers/local/worker-process.ts b/packages/server/src/server/speech/providers/local/worker-process.ts index 957ee6aaa2..2afa3d336a 100644 --- a/packages/server/src/server/speech/providers/local/worker-process.ts +++ b/packages/server/src/server/speech/providers/local/worker-process.ts @@ -3,6 +3,7 @@ import pino from "pino"; import type { StreamingTranscriptionSession } from "../../speech-provider.js"; import type { TurnDetectionSession } from "../../turn-detection-provider.js"; import { getLocalSpeechModelDir, type LocalSttModelId, type LocalTtsModelId } from "./models.js"; +import { getLocalSttEngine, getLocalTtsEngine } from "./sherpa/model-catalog.js"; import { SherpaOfflineRecognizerEngine } from "./sherpa/sherpa-offline-recognizer.js"; import { SherpaOnnxParakeetSTT } from "./sherpa/sherpa-parakeet-stt.js"; import { SherpaParakeetRealtimeTranscriptionSession } from "./sherpa/sherpa-parakeet-realtime-session.js"; @@ -85,13 +86,8 @@ function getSttEngine( const modelDir = getLocalSpeechModelDir(config.modelsDir, modelId); const created = new SherpaOfflineRecognizerEngine( { - model: { - kind: "nemo_transducer", - encoder: `${modelDir}/encoder.int8.onnx`, - decoder: `${modelDir}/decoder.int8.onnx`, - joiner: `${modelDir}/joiner.int8.onnx`, - tokens: `${modelDir}/tokens.txt`, - }, + modelDir, + engine: getLocalSttEngine(modelId), numThreads: 2, debug: 0, }, @@ -125,8 +121,9 @@ function getTtsProvider(config: LocalSpeechWorkerConfig): SherpaOnnxTTS { const modelDir = getLocalSpeechModelDir(config.modelsDir, ttsModelId(config)); const created = new SherpaOnnxTTS( { - preset: ttsModelId(config), + modelId: ttsModelId(config), modelDir, + engine: getLocalTtsEngine(ttsModelId(config)), speakerId: config.voiceTtsSpeakerId, speed: config.voiceTtsSpeed, }, diff --git a/packages/server/src/server/speech/providers/openai/config.test.ts b/packages/server/src/server/speech/providers/openai/config.test.ts index 1f3433a347..ce24ff68ef 100644 --- a/packages/server/src/server/speech/providers/openai/config.test.ts +++ b/packages/server/src/server/speech/providers/openai/config.test.ts @@ -224,6 +224,23 @@ describe("resolveOpenAiSpeechConfig", () => { expect(resolved?.tts?.voice).toBe("Aiden"); }); + test("lets STT_MODEL override the model of every feature on OpenAI", () => { + const persisted = PersistedConfigSchema.parse({ + features: { + dictation: { stt: { provider: "openai", model: "gpt-4o-mini-transcribe" } }, + voiceMode: { stt: { provider: "openai", model: "whisper-1" } }, + }, + }); + const env = { OPENAI_API_KEY: "sk-test", STT_MODEL: "gpt-4o-transcribe" } as NodeJS.ProcessEnv; + + const resolved = resolveOpenAiSpeechConfig({ env, persisted, providers: ALL_OPENAI }); + + expect(resolved?.sttModels).toEqual({ + dictation: "gpt-4o-transcribe", + voiceMode: "gpt-4o-transcribe", + }); + }); + test("accepts OpenAI's own gpt-4o-mini-tts model", () => { const persisted = PersistedConfigSchema.parse({ providers: { openai: { tts: { apiKey: "tts-key" } } }, diff --git a/packages/server/src/server/speech/providers/openai/config.ts b/packages/server/src/server/speech/providers/openai/config.ts index efb25bffaf..4e89e40ba1 100644 --- a/packages/server/src/server/speech/providers/openai/config.ts +++ b/packages/server/src/server/speech/providers/openai/config.ts @@ -7,8 +7,13 @@ import type { TTSConfig } from "./tts.js"; export const DEFAULT_OPENAI_TTS_MODEL = "tts-1"; +/** The features that transcribe through OpenAI; each picks its own model. */ +export type OpenAiSttFeature = "dictation" | "voiceMode"; + export interface OpenAiSpeechProviderConfig { - stt?: Partial & { apiKey?: string }; + /** The transcription endpoint the features share. */ + stt?: Omit, "model"> & { apiKey?: string }; + sttModels?: Partial>; tts?: Partial & { apiKey?: string }; } @@ -37,7 +42,8 @@ const OpenAiEndpointKeysSchema = z.object({ const OpenAiSttOptionsSchema = z.object({ sttConfidenceThreshold: OptionalFiniteNumberSchema, - sttModel: OptionalTrimmedStringSchema, + dictationSttModel: OptionalTrimmedStringSchema, + voiceModeSttModel: OptionalTrimmedStringSchema, }); // Voice and model are free-form so a self-hosted server can use its own ids; the @@ -85,10 +91,14 @@ function buildOpenAiSttInput(params: { env.STT_CONFIDENCE_THRESHOLD, persisted.features?.dictation?.stt?.confidenceThreshold, ]), - sttModel: firstDefined([ + // STT_MODEL is host-wide: it overrides the model of every feature that uses OpenAI. + dictationSttModel: firstDefined([ + env.STT_MODEL, + pickIfOpenAi(providers.dictationStt, persisted.features?.dictation?.stt?.model), + ]), + voiceModeSttModel: firstDefined([ env.STT_MODEL, pickIfOpenAi(providers.voiceStt, persisted.features?.voiceMode?.stt?.model), - pickIfOpenAi(providers.dictationStt, persisted.features?.dictation?.stt?.model), ]), }; } @@ -163,7 +173,7 @@ export function resolveOpenAiSpeechConfig(params: { } return { - ...(keys.sttApiKey ? { stt: buildSttConfig(keys.sttApiKey, keys.sttBaseUrl, input) } : {}), + ...(keys.sttApiKey ? buildSttConfig(keys.sttApiKey, keys.sttBaseUrl, input) : {}), ...(keys.ttsApiKey ? { tts: buildTtsConfig(keys.ttsApiKey, keys.ttsBaseUrl, input) } : {}), }; } @@ -172,15 +182,20 @@ function buildSttConfig( apiKey: string, baseUrl: string | undefined, input: Record, -): OpenAiSpeechProviderConfig["stt"] { +): Pick { const options = OpenAiSttOptionsSchema.parse(input); return { - apiKey, - ...(baseUrl ? { baseUrl } : {}), - ...(options.sttConfidenceThreshold !== undefined - ? { confidenceThreshold: options.sttConfidenceThreshold } - : {}), - ...(options.sttModel ? { model: options.sttModel } : {}), + stt: { + apiKey, + ...(baseUrl ? { baseUrl } : {}), + ...(options.sttConfidenceThreshold !== undefined + ? { confidenceThreshold: options.sttConfidenceThreshold } + : {}), + }, + sttModels: { + ...(options.dictationSttModel ? { dictation: options.dictationSttModel } : {}), + ...(options.voiceModeSttModel ? { voiceMode: options.voiceModeSttModel } : {}), + }, }; } diff --git a/packages/server/src/server/speech/providers/openai/runtime.test.ts b/packages/server/src/server/speech/providers/openai/runtime.test.ts index c947d86600..34aa35d758 100644 --- a/packages/server/src/server/speech/providers/openai/runtime.test.ts +++ b/packages/server/src/server/speech/providers/openai/runtime.test.ts @@ -1,6 +1,11 @@ +import { createServer, type Server } from "node:http"; +import type { AddressInfo } from "node:net"; import pino from "pino"; -import { describe, expect, test } from "vitest"; +import { afterEach, describe, expect, test } from "vitest"; +import { PersistedConfigSchema } from "../../../persisted-config.js"; +import type { SpeechToTextProvider } from "../../speech-provider.js"; +import { resolveOpenAiSpeechConfig } from "./config.js"; import { initializeOpenAiSpeechServices } from "./runtime.js"; import { OpenAISTT } from "./stt.js"; import { OpenAITTS } from "./tts.js"; @@ -32,3 +37,88 @@ describe("initializeOpenAiSpeechServices", () => { expect(services.ttsService).toBeInstanceOf(OpenAITTS); }); }); + +// An OpenAI-compatible transcription endpoint that records the model each request asks for. +async function startTranscriptionServer(): Promise<{ + server: Server; + baseUrl: string; + models: string[]; +}> { + const models: string[] = []; + const server = createServer((request, response) => { + let body = ""; + request.on("data", (chunk: Buffer) => { + body += chunk.toString("latin1"); + }); + request.on("end", () => { + const model = /name="model"\r\n\r\n([^\r]*)/.exec(body)?.[1]; + models.push(model ?? ""); + response.writeHead(200, { "content-type": "application/json" }); + response.end(JSON.stringify({ text: "hello" })); + }); + }); + await new Promise((resolve) => server.listen(0, "127.0.0.1", resolve)); + const { port } = server.address() as AddressInfo; + return { server, baseUrl: `http://127.0.0.1:${port}/v1`, models }; +} + +async function transcribeOnce(stt: SpeechToTextProvider | null): Promise { + if (!stt) throw new Error("STT service was not created"); + const session = stt.createSession({ logger: pino({ level: "silent" }), language: "en" }); + const transcript = new Promise((resolve, reject) => { + session.on("transcript", (event) => { + if (event.isFinal) resolve(); + }); + session.on("error", (error) => + reject(error instanceof Error ? error : new Error(String(error))), + ); + }); + await session.connect(); + session.appendPcm16(Buffer.alloc(3200)); + session.commit(); + await transcript; + session.close(); +} + +describe("OpenAI speech-to-text models", () => { + let server: Server | null = null; + + afterEach(async () => { + await new Promise((resolve) => (server ? server.close(() => resolve()) : resolve())); + server = null; + }); + + test("dictation and voice mode each transcribe with the model chosen for them", async () => { + const endpoint = await startTranscriptionServer(); + server = endpoint.server; + const persisted = PersistedConfigSchema.parse({ + features: { + dictation: { stt: { provider: "openai", model: "gpt-4o-mini-transcribe" } }, + voiceMode: { enabled: true, stt: { provider: "openai", model: "whisper-1" } }, + }, + providers: { openai: { apiKey: "sk-test", baseUrl: endpoint.baseUrl } }, + }); + const providers = { + dictationStt: { provider: "openai", explicit: true }, + voiceTurnDetection: { provider: "local", explicit: false, enabled: false }, + voiceStt: { provider: "openai", explicit: true }, + voiceTts: { provider: "local", explicit: false }, + } as const; + + const services = initializeOpenAiSpeechServices({ + providers, + openaiConfig: resolveOpenAiSpeechConfig({ env: {}, persisted, providers }), + existing: { + turnDetectionService: null, + sttService: null, + ttsService: null, + dictationSttService: null, + }, + logger: pino({ level: "silent" }), + }); + await transcribeOnce(services.dictationSttService); + await transcribeOnce(services.sttService); + + expect(endpoint.models).toEqual(["gpt-4o-mini-transcribe", "whisper-1"]); + }); +}); diff --git a/packages/server/src/server/speech/providers/openai/runtime.ts b/packages/server/src/server/speech/providers/openai/runtime.ts index a1231ddbe6..4e967e9cb4 100644 --- a/packages/server/src/server/speech/providers/openai/runtime.ts +++ b/packages/server/src/server/speech/providers/openai/runtime.ts @@ -3,7 +3,11 @@ import type { Logger } from "pino"; import type { SpeechToTextProvider, TextToSpeechProvider } from "../../speech-provider.js"; import type { RequestedSpeechProviders } from "../../speech-types.js"; import type { TurnDetectionProvider } from "../../turn-detection-provider.js"; -import { DEFAULT_OPENAI_TTS_MODEL, type OpenAiSpeechProviderConfig } from "./config.js"; +import { + DEFAULT_OPENAI_TTS_MODEL, + type OpenAiSpeechProviderConfig, + type OpenAiSttFeature, +} from "./config.js"; import { OpenAISTT } from "./stt.js"; import { OpenAITTS } from "./tts.js"; @@ -94,13 +98,18 @@ export function validateOpenAiCredentialRequirements(params: { } } -function createOpenAiStt( - apiKey: string, - openaiConfig: OpenAiSpeechProviderConfig | undefined, - logger: Logger, -): SpeechToTextProvider { - const { apiKey: _sttApiKey, ...sttConfig } = openaiConfig?.stt ?? {}; - return new OpenAISTT({ apiKey, ...sttConfig }, logger); +function createOpenAiStt(params: { + apiKey: string; + openaiConfig: OpenAiSpeechProviderConfig | undefined; + feature: OpenAiSttFeature; + logger: Logger; +}): SpeechToTextProvider { + const { apiKey: _sttApiKey, ...sttConfig } = params.openaiConfig?.stt ?? {}; + const model = params.openaiConfig?.sttModels?.[params.feature]; + return new OpenAISTT( + { apiKey: params.apiKey, ...sttConfig, ...(model ? { model } : {}) }, + params.logger, + ); } function createOpenAiTts( @@ -154,7 +163,12 @@ export function initializeOpenAiSpeechServices(params: { logger.info("OpenAI speech provider initialized"); if (needsOpenAiStt && openAiCredentials.openaiSttApiKey) { - sttService = createOpenAiStt(openAiCredentials.openaiSttApiKey, openaiConfig, logger); + sttService = createOpenAiStt({ + apiKey: openAiCredentials.openaiSttApiKey, + openaiConfig, + feature: "voiceMode", + logger, + }); } if (needsOpenAiTts && openAiCredentials.openaiTtsApiKey) { @@ -162,11 +176,12 @@ export function initializeOpenAiSpeechServices(params: { } if (needsOpenAiDictation && openAiCredentials.openaiDictationApiKey) { - dictationSttService = createOpenAiStt( - openAiCredentials.openaiDictationApiKey, + dictationSttService = createOpenAiStt({ + apiKey: openAiCredentials.openaiDictationApiKey, openaiConfig, + feature: "dictation", logger, - ); + }); } } else if (needsAnyOpenAi) { // validateOpenAiCredentialRequirements already warned about missing credentials diff --git a/packages/server/src/server/speech/speech-config-resolver.test.ts b/packages/server/src/server/speech/speech-config-resolver.test.ts index 0ee9df888a..1300dadf7e 100644 --- a/packages/server/src/server/speech/speech-config-resolver.test.ts +++ b/packages/server/src/server/speech/speech-config-resolver.test.ts @@ -6,7 +6,7 @@ import { PersistedConfigSchema } from "../persisted-config.js"; import { resolveSpeechConfig } from "./speech-config-resolver.js"; describe("resolveSpeechConfig", () => { - test("resolves local-first defaults without env overrides", () => { + test("resolves local-first defaults with voice mode off until turned on", () => { const paseoHome = "/tmp/paseo-home"; const persisted = PersistedConfigSchema.parse({}); const env = {} as NodeJS.ProcessEnv; @@ -26,17 +26,17 @@ describe("resolveSpeechConfig", () => { expect(result.speech.providers.voiceTurnDetection).toEqual({ provider: "local", explicit: false, - enabled: true, + enabled: false, }); expect(result.speech.providers.voiceStt).toEqual({ provider: "local", explicit: false, - enabled: true, + enabled: false, }); expect(result.speech.providers.voiceTts).toEqual({ provider: "local", explicit: false, - enabled: true, + enabled: false, }); expect(result.speech.local).toEqual({ modelsDir: path.join(paseoHome, "models", "local-speech"), @@ -61,6 +61,7 @@ describe("resolveSpeechConfig", () => { const persisted = PersistedConfigSchema.parse({ features: { voiceMode: { + enabled: true, turnDetection: { provider: "local" }, stt: { provider: "openai", model: "gpt-4o-transcribe" }, }, @@ -131,7 +132,7 @@ describe("resolveSpeechConfig", () => { }); expect(result.openai?.stt?.apiKey).toBe("persisted-key"); expect(result.openai?.tts?.apiKey).toBe("persisted-key"); - expect(result.openai?.stt?.model).toBe("gpt-4o-transcribe"); + expect(result.openai?.sttModels).toEqual({ voiceMode: "gpt-4o-transcribe" }); }); test("resolves STT language from env, settings, and voice-to-dictation fallback", () => { diff --git a/packages/server/src/server/speech/speech-config-resolver.ts b/packages/server/src/server/speech/speech-config-resolver.ts index 95c78707a5..ae0c299aa1 100644 --- a/packages/server/src/server/speech/speech-config-resolver.ts +++ b/packages/server/src/server/speech/speech-config-resolver.ts @@ -43,8 +43,13 @@ const RequestedSpeechProvidersSchema = z.object({ voiceTts: OptionalSpeechProviderSchema.default("local"), }); -function resolveOptionalBooleanFlag(value: unknown): boolean { - return OptionalBooleanFlagSchema.parse(value) ?? true; +// Dictation is on unless turned off. Voice mode is opt-in: it downloads its own models and +// runs a hidden agent, so a host only pays for it once someone turns it on. +export const DICTATION_ENABLED_BY_DEFAULT = true; +export const VOICE_MODE_ENABLED_BY_DEFAULT = false; + +function resolveOptionalBooleanFlag(value: unknown, fallback: boolean): boolean { + return OptionalBooleanFlagSchema.parse(value) ?? fallback; } interface FeatureProviderInputs { @@ -70,6 +75,7 @@ function buildFeatureProviderInputs(params: { params.env.PASEO_VOICE_MODE_ENABLED, params.persisted.features?.voiceMode?.enabled, ]), + VOICE_MODE_ENABLED_BY_DEFAULT, ); return { dictationStt: { @@ -82,6 +88,7 @@ function buildFeatureProviderInputs(params: { params.env.PASEO_DICTATION_ENABLED, params.persisted.features?.dictation?.enabled, ]), + DICTATION_ENABLED_BY_DEFAULT, ), }, voiceTurnDetection: { @@ -173,6 +180,7 @@ export function resolveSpeechConfig(params: { openai, speech: { providers, + localModelsDir: local.modelsDir, sttLanguages: local.sttLanguages, ...(local.local ? { local: local.local } : {}), }, diff --git a/packages/server/src/server/speech/speech-runtime.test.ts b/packages/server/src/server/speech/speech-runtime.test.ts index c9c21b3e89..cd3580e64a 100644 --- a/packages/server/src/server/speech/speech-runtime.test.ts +++ b/packages/server/src/server/speech/speech-runtime.test.ts @@ -2,13 +2,14 @@ import { beforeEach, describe, expect, it, vi } from "vitest"; import pino from "pino"; import type { PaseoSpeechConfig } from "../bootstrap.js"; +import type { LocalSpeechModelId } from "./providers/local/models.js"; import type { InitializedLocalSpeech } from "./providers/local/runtime.js"; import type { SpeechToTextProvider, TextToSpeechProvider } from "./speech-provider.js"; import type { TurnDetectionProvider } from "./turn-detection-provider.js"; -import { createSpeechService } from "./speech-runtime.js"; +import { createSpeechService, LocalSpeechModelInUseError } from "./speech-runtime.js"; -const { ensureLocalSpeechModelsMock, initializeLocalSpeechServicesMock } = vi.hoisted(() => ({ - ensureLocalSpeechModelsMock: vi.fn(async () => ({})), +const { ensureSherpaOnnxModelMock, initializeLocalSpeechServicesMock } = vi.hoisted(() => ({ + ensureSherpaOnnxModelMock: vi.fn(async (_options: { signal?: AbortSignal }) => ""), initializeLocalSpeechServicesMock: vi.fn<(args: unknown) => Promise>(), })); @@ -34,9 +35,10 @@ vi.mock("./providers/openai/runtime.js", () => ({ validateOpenAiCredentialRequirements: () => {}, })); -vi.mock("./providers/local/models.js", () => ({ - ensureLocalSpeechModels: ensureLocalSpeechModelsMock, - listMissingLocalSpeechModels: vi.fn(async ({ modelIds }: { modelIds: string[] }) => modelIds), +vi.mock("./providers/local/sherpa/model-downloader.js", () => ({ + ensureSherpaOnnxModel: ensureSherpaOnnxModelMock, + isSherpaOnnxModelInstalled: vi.fn(async () => false), + deleteSherpaOnnxModel: vi.fn(async () => {}), })); function createStubStt(id: string): SpeechToTextProvider { @@ -75,6 +77,7 @@ function rejectWhenAborted(signal?: AbortSignal): Promise { function createSpeechConfig(providers: PaseoSpeechConfig["providers"]): PaseoSpeechConfig { return { providers, + localModelsDir: "/tmp/missing-local-speech-models", sttLanguages: { dictation: "en", voice: "en", @@ -84,8 +87,8 @@ function createSpeechConfig(providers: PaseoSpeechConfig["providers"]): PaseoSpe describe("createSpeechService readiness", () => { beforeEach(() => { - ensureLocalSpeechModelsMock.mockReset(); - ensureLocalSpeechModelsMock.mockResolvedValue({}); + ensureSherpaOnnxModelMock.mockReset(); + ensureSherpaOnnxModelMock.mockResolvedValue(""); initializeLocalSpeechServicesMock.mockReset(); }); @@ -169,7 +172,7 @@ describe("createSpeechService readiness", () => { it("aborts and joins an in-flight model download when stopped", async () => { let downloadSignal: AbortSignal | undefined; - ensureLocalSpeechModelsMock.mockImplementation(({ signal }: { signal?: AbortSignal }) => { + ensureSherpaOnnxModelMock.mockImplementation(({ signal }: { signal?: AbortSignal }) => { downloadSignal = signal; return rejectWhenAborted(signal); }); @@ -201,10 +204,104 @@ describe("createSpeechService readiness", () => { }); runtime.start(); await runtime.ready; - expect(ensureLocalSpeechModelsMock).toHaveBeenCalledOnce(); + expect(ensureSherpaOnnxModelMock).toHaveBeenCalledOnce(); await runtime.stop(); expect(downloadSignal?.aborted).toBe(true); }); }); + +describe("createSpeechService settings changes", () => { + const dictationOnly = createSpeechConfig({ + dictationStt: { provider: "local", enabled: true, explicit: true }, + voiceTurnDetection: { provider: "local", enabled: false, explicit: true }, + voiceStt: { provider: "local", enabled: false, explicit: true }, + voiceTts: { provider: "local", enabled: false, explicit: true }, + }); + const nothingEnabled = createSpeechConfig({ + dictationStt: { provider: "local", enabled: false, explicit: true }, + voiceTurnDetection: { provider: "local", enabled: false, explicit: true }, + voiceStt: { provider: "local", enabled: false, explicit: true }, + voiceTts: { provider: "local", enabled: false, explicit: true }, + }); + + function localSpeech(input: { + dictationStt: SpeechToTextProvider | null; + defaultModelIds: LocalSpeechModelId[]; + }): InitializedLocalSpeech { + return { + turnDetectionService: null, + sttService: null, + ttsService: null, + dictationSttService: input.dictationStt, + localVoiceTtsProvider: null, + localModelConfig: { + modelsDir: "/tmp/missing-local-speech-models", + defaultModelIds: input.defaultModelIds, + }, + availability: { configured: true, modelsDir: "/tmp/missing-local-speech-models" }, + cleanup: () => {}, + }; + } + + beforeEach(() => { + ensureSherpaOnnxModelMock.mockReset(); + ensureSherpaOnnxModelMock.mockResolvedValue(""); + initializeLocalSpeechServicesMock.mockReset(); + }); + + it("applies new providers without a restart", async () => { + initializeLocalSpeechServicesMock.mockResolvedValue( + localSpeech({ dictationStt: createStubStt("dictation-local"), defaultModelIds: [] }), + ); + const runtime = createSpeechService({ + logger: pino({ level: "silent" }), + speechConfig: dictationOnly, + }); + runtime.start(); + await runtime.ready; + expect(runtime.getReadiness().dictation.available).toBe(true); + + const disabled = new Promise((resolve) => { + runtime.onReadinessChange((snapshot) => { + if (snapshot.dictation.reasonCode === "disabled") resolve(); + }); + }); + initializeLocalSpeechServicesMock.mockResolvedValue( + localSpeech({ dictationStt: null, defaultModelIds: [] }), + ); + runtime.reconfigure({ speechConfig: nothingEnabled }); + await disabled; + + expect(runtime.resolveDictationStt()).toBeNull(); + await runtime.stop(); + }); + + it("refuses to delete a model an enabled feature uses", async () => { + initializeLocalSpeechServicesMock.mockResolvedValue( + localSpeech({ + dictationStt: createStubStt("dictation-local"), + defaultModelIds: ["parakeet-tdt-0.6b-v2-int8"], + }), + ); + ensureSherpaOnnxModelMock.mockImplementation(({ signal }) => rejectWhenAborted(signal)); + const runtime = createSpeechService({ + logger: pino({ level: "silent" }), + speechConfig: dictationOnly, + }); + runtime.start(); + await runtime.ready; + + const models = await runtime.listLocalModels(); + expect(models.filter((model) => model.inUse).map((model) => model.spec.id)).toEqual([ + "parakeet-tdt-0.6b-v2-int8", + ]); + await expect(runtime.deleteLocalModel("parakeet-tdt-0.6b-v2-int8")).rejects.toBeInstanceOf( + LocalSpeechModelInUseError, + ); + await expect(runtime.deleteLocalModel("kokoro-int8-en-v0_19")).resolves.toBeUndefined(); + + await runtime.stop(); + }); +}); diff --git a/packages/server/src/server/speech/speech-runtime.ts b/packages/server/src/server/speech/speech-runtime.ts index 932f87c077..270b13b03c 100644 --- a/packages/server/src/server/speech/speech-runtime.ts +++ b/packages/server/src/server/speech/speech-runtime.ts @@ -2,7 +2,10 @@ import type { Logger } from "pino"; import type { PaseoOpenAIConfig, PaseoSpeechConfig } from "../bootstrap.js"; import type { LocalSpeechModelId } from "./providers/local/config.js"; -import { ensureLocalSpeechModels, listMissingLocalSpeechModels } from "./providers/local/models.js"; +import { + LocalSpeechModelLibrary, + type LocalSpeechModelEntry, +} from "./providers/local/model-library.js"; import { initializeLocalSpeechServices } from "./providers/local/runtime.js"; import { getOpenAiSpeechAvailability, @@ -92,7 +95,7 @@ function buildRealtimeVoiceReadiness(params: { enabled: false, available: false, reasonCode: "disabled", - message: "Realtime voice is disabled in daemon config.", + message: "Voice mode is off. Turn it on in Settings > Voice.", retryable: false, missingModelIds: [], }; @@ -147,7 +150,7 @@ function buildDictationReadiness(params: { enabled: false, available: false, reasonCode: "disabled", - message: "Dictation is disabled in daemon config.", + message: "Dictation is off. Turn it on in Settings > Voice.", retryable: false, missingModelIds: [], }; @@ -292,6 +295,23 @@ function resolveEffectiveProviderIds(params: { }; } +export interface LocalSpeechModelStatus extends LocalSpeechModelEntry { + // Selected by an enabled feature, so the runtime keeps it installed. + inUse: boolean; +} + +export class LocalSpeechModelInUseError extends Error { + constructor(public readonly modelId: LocalSpeechModelId) { + super(`${modelId} is selected for dictation or voice mode. Choose another model first.`); + this.name = "LocalSpeechModelInUseError"; + } +} + +export interface SpeechServiceConfig { + openaiConfig?: PaseoOpenAIConfig; + speechConfig?: PaseoSpeechConfig; +} + export interface SpeechService { resolveStt: () => SpeechToTextProvider | null; resolveSttLanguage: () => string; @@ -301,37 +321,49 @@ export interface SpeechService { resolveDictationSttLanguage: () => string; getReadiness: () => SpeechReadinessSnapshot; onReadinessChange: (listener: (snapshot: SpeechReadinessSnapshot) => void) => () => void; + // Applies new provider/model settings without restarting the daemon. + reconfigure: (config: SpeechServiceConfig) => void; + listLocalModels: () => Promise; + downloadLocalModel: (modelId: LocalSpeechModelId) => void; + deleteLocalModel: (modelId: LocalSpeechModelId) => Promise; start: () => void; stop: () => Promise; ready: Promise; } -export function createSpeechService(params: { - logger: Logger; - openaiConfig?: PaseoOpenAIConfig; - speechConfig?: PaseoSpeechConfig; -}): SpeechService { +export function createSpeechService( + params: SpeechServiceConfig & { + logger: Logger; + }, +): SpeechService { const logger = params.logger.child({ module: "speech-runtime" }); - const speechConfig = params.speechConfig ?? null; - const openaiConfig = params.openaiConfig; - const providers = resolveRequestedSpeechProviders(speechConfig); - const requestedProviders = describeRequestedProviders(providers); + let speechConfig = params.speechConfig ?? null; + let openaiConfig = params.openaiConfig; + let providers = resolveRequestedSpeechProviders(speechConfig); + let requestedProviders = describeRequestedProviders(providers); + // The models directory is fixed for the daemon's lifetime; settings changes only pick + // which of its models are used. + const modelLibrary = speechConfig + ? new LocalSpeechModelLibrary(speechConfig.localModelsDir, logger) + : null; - validateOpenAiCredentialRequirements({ - providers, - openaiConfig, - logger, - }); - - logger.info( - { - requestedProviders, - availability: { - openai: getOpenAiSpeechAvailability(openaiConfig), + const logReconciliationStart = (): void => { + validateOpenAiCredentialRequirements({ + providers, + openaiConfig, + logger, + }); + logger.info( + { + requestedProviders, + availability: { + openai: getOpenAiSpeechAvailability(openaiConfig), + }, }, - }, - "Speech provider reconciliation started", - ); + "Speech provider reconciliation started", + ); + }; + logReconciliationStart(); let sttService: SpeechToTextProvider | null = null; let ttsService: TextToSpeechProvider | null = null; @@ -346,12 +378,11 @@ export function createSpeechService(params: { let missingLocalModelIds: LocalSpeechModelId[] = []; let backgroundDownloadInProgress = false; - let backgroundDownloadAbortController: AbortController | null = null; let backgroundDownloadPromise: Promise | null = null; let backgroundDownloadError: string | null = null; let stopped = false; let monitorTimeout: ReturnType | null = null; - let reconcileInFlight: Promise | null = null; + let reconcileQueue: Promise = Promise.resolve(); const readinessListeners = new Set<(snapshot: SpeechReadinessSnapshot) => void>(); let lastReadinessFingerprint: string | null = null; let lastPublishedReadinessSnapshot: SpeechReadinessSnapshot | null = null; @@ -445,12 +476,10 @@ export function createSpeechService(params: { }; const refreshMissingLocalModels = async (): Promise => { - missingLocalModelIds = localModelConfig - ? await listMissingLocalSpeechModels({ - modelsDir: localModelConfig.modelsDir, - modelIds: localModelConfig.defaultModelIds, - }) - : []; + missingLocalModelIds = + localModelConfig && modelLibrary + ? await modelLibrary.listMissing(localModelConfig.defaultModelIds) + : []; }; const reconcileServices = async (): Promise => { @@ -520,16 +549,12 @@ export function createSpeechService(params: { } }; + // Reconciles run one after another, and each starts after it was requested, so a + // reconfigure that lands mid-reconcile is applied instead of joining the stale run. const runReconcile = async (): Promise => { - if (reconcileInFlight) { - await reconcileInFlight; - publishReadinessIfChanged(); - return; - } - reconcileInFlight = reconcileServices().finally(() => { - reconcileInFlight = null; - }); - await reconcileInFlight; + const next = reconcileQueue.catch(() => undefined).then(reconcileServices); + reconcileQueue = next; + await next; publishReadinessIfChanged(); }; @@ -544,12 +569,12 @@ export function createSpeechService(params: { }; const startBackgroundDownload = (): void => { - if (stopped || backgroundDownloadInProgress) { + if (stopped || backgroundDownloadInProgress || !modelLibrary) { return; } - const modelsDir = localModelConfig?.modelsDir ?? null; + const modelsDir = modelLibrary.getModelsDir(); const modelIds = [...missingLocalModelIds]; - if (!modelsDir || modelIds.length === 0) { + if (modelIds.length === 0) { return; } @@ -565,16 +590,9 @@ export function createSpeechService(params: { "Starting background download for missing local speech models", ); - const abortController = new AbortController(); - backgroundDownloadAbortController = abortController; backgroundDownloadPromise = (async () => { try { - await ensureLocalSpeechModels({ - modelsDir, - modelIds, - logger, - signal: abortController.signal, - }); + await Promise.all(modelIds.map((modelId) => modelLibrary.download(modelId))); await runReconcile(); backgroundDownloadError = null; } catch (error) { @@ -588,7 +606,6 @@ export function createSpeechService(params: { "Background local speech model download failed", ); } finally { - backgroundDownloadAbortController = null; backgroundDownloadInProgress = false; await refreshMissingLocalModels().catch((error) => { logger.warn({ err: error }, "Failed to refresh local speech model status after download"); @@ -630,6 +647,62 @@ export function createSpeechService(params: { } }; + const reconcileAndFetchMissingModels = async (): Promise => { + await runReconcile(); + const snapshot = computeReadinessSnapshot(); + if (snapshot.voiceFeature.enabled && !snapshot.voiceFeature.available) { + if (missingLocalModelIds.length > 0) { + startBackgroundDownload(); + } + scheduleMonitor(); + } + }; + + const reconfigure = (next: SpeechServiceConfig): void => { + speechConfig = next.speechConfig ?? null; + openaiConfig = next.openaiConfig; + providers = resolveRequestedSpeechProviders(speechConfig); + requestedProviders = describeRequestedProviders(providers); + // A failure for the previous selection must not block downloading the new one. + backgroundDownloadError = null; + logReconciliationStart(); + if (!started || stopped) { + return; + } + void reconcileAndFetchMissingModels().catch((error) => { + logger.error({ err: error }, "Speech runtime failed to apply new configuration"); + }); + }; + + const requireModelLibrary = (): LocalSpeechModelLibrary => { + if (!modelLibrary) { + throw new Error("Local speech models are unavailable on this daemon"); + } + return modelLibrary; + }; + + const listLocalModels = async (): Promise => { + const inUse = new Set(localModelConfig?.defaultModelIds ?? []); + const entries = await requireModelLibrary().list(); + return entries.map(({ spec, state }) => ({ spec, state, inUse: inUse.has(spec.id) })); + }; + + const downloadLocalModel = (modelId: LocalSpeechModelId): void => { + void requireModelLibrary() + .download(modelId) + .then(() => runMonitorTick()) + .catch((error) => { + logger.warn({ err: error, modelId }, "Local speech model download failed"); + }); + }; + + const deleteLocalModel = async (modelId: LocalSpeechModelId): Promise => { + if (localModelConfig?.defaultModelIds.includes(modelId)) { + throw new LocalSpeechModelInUseError(modelId); + } + await requireModelLibrary().delete(modelId); + }; + const start = (): void => { if (started || stopped) { return; @@ -637,14 +710,7 @@ export function createSpeechService(params: { started = true; void (async () => { try { - await runReconcile(); - const snapshot = computeReadinessSnapshot(); - if (snapshot.voiceFeature.enabled && !snapshot.voiceFeature.available) { - if (missingLocalModelIds.length > 0) { - startBackgroundDownload(); - } - scheduleMonitor(); - } + await reconcileAndFetchMissingModels(); if (!readySettled) { readySettled = true; resolveReady(); @@ -666,7 +732,7 @@ export function createSpeechService(params: { monitorTimeout = null; } localCleanup(); - backgroundDownloadAbortController?.abort(); + await modelLibrary?.stop(); await backgroundDownloadPromise?.catch(() => undefined); backgroundDownloadPromise = null; }; @@ -680,6 +746,10 @@ export function createSpeechService(params: { resolveDictationSttLanguage: () => speechConfig?.sttLanguages?.dictation ?? "en", getReadiness: () => lastPublishedReadinessSnapshot ?? computeReadinessSnapshot(), onReadinessChange: subscribeSpeechReadiness, + reconfigure, + listLocalModels, + downloadLocalModel, + deleteLocalModel, start, stop, ready, diff --git a/packages/server/src/server/speech/speech-settings.ts b/packages/server/src/server/speech/speech-settings.ts new file mode 100644 index 0000000000..a04a8bed59 --- /dev/null +++ b/packages/server/src/server/speech/speech-settings.ts @@ -0,0 +1,211 @@ +import type { SpeechSettings, SpeechSettingsPatch } from "@getpaseo/protocol/speech/rpc-schemas"; + +import type { PersistedConfig } from "../persisted-config.js"; +import { + DICTATION_ENABLED_BY_DEFAULT, + VOICE_MODE_ENABLED_BY_DEFAULT, + resolveSpeechConfig, +} from "./speech-config-resolver.js"; +import type { SpeechService } from "./speech-runtime.js"; +import { LocalSttModelIdSchema, LocalTtsModelIdSchema } from "./providers/local/models.js"; +import { SpeechProviderIdSchema, type SpeechProviderId } from "./speech-types.js"; + +// The `speech` section of the mutable daemon config is a view over three persisted +// blocks: features.dictation, features.voiceMode and providers.openai. This module owns +// both directions of that mapping. + +type PersistedFeatures = NonNullable; +type PersistedDictation = NonNullable; +type PersistedVoiceMode = NonNullable; +type PersistedStt = NonNullable; +type PersistedTts = NonNullable; +type PersistedOpenAi = NonNullable["openai"]>; +type SttPatch = NonNullable["stt"]>; +type TtsPatch = NonNullable["tts"]>; + +const DEFAULT_SPEECH_PROVIDER = "local"; + +// The daemon config store derives the speech view through this owner (the OpenAI key is +// write-only) and calls apply when the persisted speech blocks change. +export interface SpeechSettingsOwner { + read(persisted: PersistedConfig): SpeechSettings; + apply(persisted: PersistedConfig): void; +} + +export function createSpeechSettingsOwner(params: { + paseoHome: string; + env: NodeJS.ProcessEnv; + speechService: Pick; +}): SpeechSettingsOwner { + const { paseoHome, env, speechService } = params; + return { + read: (persisted) => readSpeechSettings(persisted, env), + apply: (persisted) => { + const resolved = resolveSpeechConfig({ paseoHome, env, persisted }); + speechService.reconfigure({ openaiConfig: resolved.openai, speechConfig: resolved.speech }); + }, + }; +} + +export function readSpeechSettings( + persisted: PersistedConfig, + env: NodeJS.ProcessEnv, +): SpeechSettings { + const dictation = persisted.features?.dictation; + const voiceMode = persisted.features?.voiceMode; + return { + dictation: { + enabled: dictation?.enabled ?? DICTATION_ENABLED_BY_DEFAULT, + stt: readStt(dictation?.stt), + }, + voiceMode: { + enabled: voiceMode?.enabled ?? VOICE_MODE_ENABLED_BY_DEFAULT, + stt: readStt(voiceMode?.stt), + tts: readTts(voiceMode?.tts), + }, + openai: readOpenAi(persisted.providers?.openai, env), + }; +} + +function readTts(tts: PersistedTts | undefined): SpeechSettings["voiceMode"]["tts"] { + return { + provider: tts?.provider ?? DEFAULT_SPEECH_PROVIDER, + ...(tts?.model ? { model: tts.model } : {}), + ...(tts?.voice ? { voice: tts.voice } : {}), + }; +} + +function readOpenAi( + openai: PersistedOpenAi | undefined, + env: NodeJS.ProcessEnv, +): SpeechSettings["openai"] { + const keys = [ + openai?.apiKey, + openai?.stt?.apiKey, + openai?.tts?.apiKey, + env.OPENAI_API_KEY, + env.OPENAI_STT_API_KEY, + env.OPENAI_TTS_API_KEY, + ]; + return { + apiKeyConfigured: keys.some((key) => Boolean(key?.trim())), + ...(openai?.baseUrl ? { baseUrl: openai.baseUrl } : {}), + }; +} + +function readStt(stt: PersistedStt | undefined): SpeechSettings["dictation"]["stt"] { + return { + provider: stt?.provider ?? DEFAULT_SPEECH_PROVIDER, + ...(stt?.model ? { model: stt.model } : {}), + ...(stt?.language ? { language: stt.language } : {}), + }; +} + +export function mergeSpeechSettingsPatch( + persisted: PersistedConfig, + patch: SpeechSettingsPatch, +): PersistedConfig { + const features: PersistedFeatures = { ...persisted.features }; + if (patch.dictation) { + const current = features.dictation; + const stt = patch.dictation.stt ? mergeStt(current?.stt, patch.dictation.stt) : current?.stt; + features.dictation = withoutUndefined({ + ...current, + enabled: patch.dictation.enabled ?? current?.enabled, + stt, + }); + } + if (patch.voiceMode) { + const current = features.voiceMode; + features.voiceMode = withoutUndefined({ + ...current, + enabled: patch.voiceMode.enabled ?? current?.enabled, + stt: patch.voiceMode.stt ? mergeStt(current?.stt, patch.voiceMode.stt) : current?.stt, + tts: patch.voiceMode.tts ? mergeTts(current?.tts, patch.voiceMode.tts) : current?.tts, + }); + } + + const providers = { ...persisted.providers }; + if (patch.openai) { + providers.openai = mergeOpenAi(providers.openai, patch.openai); + } + + return { ...persisted, features, providers }; +} + +function mergeStt(current: PersistedStt | undefined, patch: SttPatch): PersistedStt { + const provider = parseProvider(patch.provider) ?? current?.provider; + const providerChanged = patch.provider !== undefined && patch.provider !== current?.provider; + const model = resolvePatchedValue(providerChanged ? undefined : current?.model, patch.model); + assertLocalModelKind(provider, model, "stt"); + return withoutUndefined({ + ...current, + provider, + model, + language: resolvePatchedValue(current?.language, patch.language), + }); +} + +function mergeTts(current: PersistedTts | undefined, patch: TtsPatch): PersistedTts { + const provider = parseProvider(patch.provider) ?? current?.provider; + const providerChanged = patch.provider !== undefined && patch.provider !== current?.provider; + const model = resolvePatchedValue(providerChanged ? undefined : current?.model, patch.model); + assertLocalModelKind(provider, model, "tts"); + return withoutUndefined({ + ...current, + provider, + model, + voice: resolvePatchedValue(providerChanged ? undefined : current?.voice, patch.voice), + }); +} + +function mergeOpenAi( + current: PersistedOpenAi | undefined, + patch: NonNullable, +): PersistedOpenAi | undefined { + const next = withoutUndefined({ + ...current, + apiKey: resolvePatchedValue(current?.apiKey, trimToNull(patch.apiKey)), + baseUrl: resolvePatchedValue(current?.baseUrl, trimToNull(patch.baseUrl)), + }); + return Object.keys(next).length > 0 ? next : undefined; +} + +// A local model id that is not in the catalog would stop the daemon from resolving its +// speech config at the next start, so reject it before it is written. +function assertLocalModelKind( + provider: string | undefined, + model: string | undefined, + kind: "stt" | "tts", +): void { + if ((provider ?? DEFAULT_SPEECH_PROVIDER) !== "local" || model === undefined) { + return; + } + const schema = kind === "stt" ? LocalSttModelIdSchema : LocalTtsModelIdSchema; + if (!schema.safeParse(model).success) { + throw new Error(`Unknown local ${kind === "stt" ? "speech-to-text" : "voice"} model: ${model}`); + } +} + +function parseProvider(provider: string | undefined): SpeechProviderId | undefined { + return provider === undefined ? undefined : SpeechProviderIdSchema.parse(provider); +} + +// undefined keeps the current value; null clears it. +function resolvePatchedValue( + current: string | undefined, + patch: string | null | undefined, +): string | undefined { + if (patch === undefined) return current; + return patch ?? undefined; +} + +function trimToNull(value: string | null | undefined): string | null | undefined { + if (typeof value !== "string") return value; + const trimmed = value.trim(); + return trimmed.length > 0 ? trimmed : null; +} + +function withoutUndefined(value: T): T { + return Object.fromEntries(Object.entries(value).filter(([, entry]) => entry !== undefined)) as T; +} diff --git a/packages/server/src/server/test-utils/paseo-daemon.ts b/packages/server/src/server/test-utils/paseo-daemon.ts index 91b125328c..fef9d76014 100644 --- a/packages/server/src/server/test-utils/paseo-daemon.ts +++ b/packages/server/src/server/test-utils/paseo-daemon.ts @@ -38,7 +38,8 @@ interface TestPaseoDaemonOptions { staticDir?: string; cleanup?: boolean; openai?: PaseoOpenAIConfig; - speech?: PaseoSpeechConfig; + // localModelsDir defaults to the local provider's modelsDir, then to a dir under paseoHome. + speech?: Omit & { localModelsDir?: string }; voiceLlmProvider?: PaseoDaemonConfig["voiceLlmProvider"]; voiceLlmProviderExplicit?: boolean; voiceLlmModel?: string | null; @@ -207,7 +208,15 @@ async function prepareTestDaemonConfig( webUi: options.webUi, trustedProxies: options.trustedProxies, openai: options.openai, - speech: options.speech, + speech: options.speech + ? { + ...options.speech, + localModelsDir: + options.speech.localModelsDir ?? + options.speech.local?.modelsDir ?? + path.join(paseoHome, "models", "local-speech"), + } + : undefined, voiceLlmProvider: options.voiceLlmProvider ?? null, voiceLlmProviderExplicit: options.voiceLlmProviderExplicit ?? false, voiceLlmModel: options.voiceLlmModel ?? null, diff --git a/packages/server/src/server/websocket-server.ts b/packages/server/src/server/websocket-server.ts index e833a07ce3..65111db5f3 100644 --- a/packages/server/src/server/websocket-server.ts +++ b/packages/server/src/server/websocket-server.ts @@ -1500,6 +1500,7 @@ export class VoiceAssistantWebSocketServer { stt: () => this.speech?.resolveStt() ?? null, sttLanguage: this.speech?.resolveSttLanguage() ?? "en", tts: () => this.speech?.resolveTts() ?? null, + speechModels: this.speech, terminalManager: this.terminalManager, providerSnapshotManager: this.providerSnapshotManager, hubExecutionAgents: options.hubExecutionAgents, @@ -1831,6 +1832,8 @@ export class VoiceAssistantWebSocketServer { ...(this.advertiseDaemonStatusRpc ? { daemonStatusRpc: true } : {}), // COMPAT(daemonConfigReload): added in v0.4.0, remove gate after 2027-02-14. daemonConfigReload: true, + // COMPAT(speechSettings): added in v0.12.0, remove gate after 2027-04-18. + speechSettings: true, // COMPAT(relayConfig): added in v0.2.6, remove gate after 2027-01-31. ...(this.advertiseRelayConfig ? { relayConfig: true } : {}), // COMPAT(pushTokenRevocation): added in v0.3.2, remove gate after 2027-02-10. diff --git a/public-docs/voice.md b/public-docs/voice.md index 79022c15ad..00fe21ae62 100644 --- a/public-docs/voice.md +++ b/public-docs/voice.md @@ -27,16 +27,13 @@ This keeps credentials and execution in your environment and avoids introducing Local speech defaults to model IDs `parakeet-tdt-0.6b-v2-int8` (STT) and `kokoro-en-v0_19` (TTS, speaker 0 / voice 00). -Missing models are downloaded at daemon startup into `$PASEO_HOME/models/local-speech`. Downloads happen only for missing files. +Dictation is on by default. Voice mode is off until you turn it on in **Settings > Voice** or set `features.voiceMode.enabled: true`. + +Missing models for enabled features are downloaded at daemon startup into `$PASEO_HOME/models/local-speech`. Downloads happen only for missing files. **Settings > Voice** lists every local model with its size and languages, and lets you download or delete each one. ### Local STT models and language support -| Model ID | Languages | -| --------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| `parakeet-tdt-0.6b-v2-int8` | English only (default). Includes punctuation and capitalization. | -| `parakeet-tdt-0.6b-v3-int8` | 25 European languages, auto-detected: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, Ukrainian. | - -**To use a non-English language, switch the local STT model to `parakeet-tdt-0.6b-v3-int8`.** v3 detects the spoken language automatically — there is no per-language setting for it. The `language` field below does **not** steer the local Parakeet model (v2 is English-only, v3 auto-detects); it only applies to the OpenAI STT provider. +**Settings > Voice** lists every local STT model with the languages it supports. The default, `parakeet-tdt-0.6b-v2-int8`, is English-only. For other languages, pick a multilingual model such as `parakeet-tdt-0.6b-v3-int8` (25 European languages) or `whisper-base` (99 languages). Local models detect the spoken language themselves; the `language` field below only applies to the OpenAI STT provider. ```json { @@ -59,7 +56,7 @@ Missing models are downloaded at daemon startup into `$PASEO_HOME/models/local-s } ``` -For multilingual local dictation, set the model to v3 — it auto-detects the language, so no `language` field is needed: +For multilingual local dictation, set a multilingual model. No `language` field is needed: ```json { @@ -72,7 +69,7 @@ For multilingual local dictation, set the model to v3 — it auto-detects the la } ``` -The `language` field applies only to the OpenAI STT provider: set `features.dictation.stt.language` for dictation and `features.voiceMode.stt.language` for voice mode. If voice language is omitted, Paseo uses the dictation language before falling back to `en`. It has no effect on the local Parakeet models. +The `language` field applies only to the OpenAI STT provider: set `features.dictation.stt.language` for dictation and `features.voiceMode.stt.language` for voice mode. If voice language is omitted, Paseo uses the dictation language before falling back to `en`. It has no effect on local models. ## OpenAI Voice Option