dsh-live-voice 0.3.0 → 0.3.2

package/lib/server.js

@@ -7,17 +7,17 @@ import { randomUUID as randomUUID2 } from "node:crypto";
 // src/modules/core/settings.ts
 var voiceDetectionPresets = Object.freeze({
   short: Object.freeze({
-    silenceMs: 900,
+    silenceMs: 500,
     label: "Short",
     description: "Send quickly after a short pause."
   }),
   natural: Object.freeze({
-    silenceMs: 1500,
+    silenceMs: 1e3,
     label: "Natural",
     description: "Allow normal pauses between phrases."
   }),
   long: Object.freeze({
-    silenceMs: 2200,
+    silenceMs: 2e3,
     label: "Long",
     description: "Wait through longer thinking pauses."
   })
@@ -43,7 +43,8 @@ var defaultSettings = Object.freeze({
   recognitionEngine: "browser",
   recognitionProcessLocally: true,
   recognitionAutoInstall: true,
-  voiceDetectionPreset: "natural",
+  voiceDetectionPreset: "short",
+  voiceDetectionCustomSilenceMs: 1e3,
   recognitionMaxUtteranceSeconds: 60,
   microphoneEnabled: true,
   holdToTalkEnabled: true,
@@ -76,6 +77,73 @@ var defaultSettings = Object.freeze({
   rate: 1,
   segmentGapMs: 400
 });
+var normalizeCommandPhrases = (value, fallback) => typeof value === "string" && value.length <= 1e3 && !value.includes("\0") ? value.split(/[,\r\n]+/).map((phrase) => phrase.trim()).filter(Boolean).slice(0, 20).join(", ") : fallback;
+var customSilenceMinMs = 100;
+var customSilenceMaxMs = 1e4;
+var normalizeCustomSilenceMs = (value) => Number.isInteger(value) && value >= customSilenceMinMs && value <= customSilenceMaxMs ? value : defaultSettings.voiceDetectionCustomSilenceMs;
+function normalizeSettings(value) {
+  const source = value && typeof value === "object" && !Array.isArray(value) ? value : {};
+  return {
+    engine: ["browser", "say", "qwen-http"].includes(source.engine) ? source.engine : defaultSettings.engine,
+    recognitionEngine: ["browser", "whisper-http", "qwen-http"].includes(source.recognitionEngine) ? source.recognitionEngine : defaultSettings.recognitionEngine,
+    recognitionProcessLocally: typeof source.recognitionProcessLocally === "boolean" ? source.recognitionProcessLocally : defaultSettings.recognitionProcessLocally,
+    recognitionAutoInstall: typeof source.recognitionAutoInstall === "boolean" ? source.recognitionAutoInstall : defaultSettings.recognitionAutoInstall,
+    voiceDetectionPreset: source.voiceDetectionPreset === "custom" || Object.hasOwn(voiceDetectionPresets, source.voiceDetectionPreset) ? source.voiceDetectionPreset : defaultSettings.voiceDetectionPreset,
+    voiceDetectionCustomSilenceMs: normalizeCustomSilenceMs(source.voiceDetectionCustomSilenceMs),
+    recognitionMaxUtteranceSeconds: Number.isInteger(source.recognitionMaxUtteranceSeconds) && source.recognitionMaxUtteranceSeconds >= 10 && source.recognitionMaxUtteranceSeconds <= 300 ? source.recognitionMaxUtteranceSeconds : defaultSettings.recognitionMaxUtteranceSeconds,
+    microphoneEnabled: typeof source.microphoneEnabled === "boolean" ? source.microphoneEnabled : defaultSettings.microphoneEnabled,
+    holdToTalkEnabled: typeof source.holdToTalkEnabled === "boolean" ? source.holdToTalkEnabled : defaultSettings.holdToTalkEnabled,
+    announceAssistantMessages: typeof source.announceAssistantMessages === "boolean" ? source.announceAssistantMessages : defaultSettings.announceAssistantMessages,
+    agentVoiceContextEnabled: typeof source.agentVoiceContextEnabled === "boolean" ? source.agentVoiceContextEnabled : defaultSettings.agentVoiceContextEnabled,
+    agentVoiceContext: typeof source.agentVoiceContext === "string" && source.agentVoiceContext.length <= 4e3 && !source.agentVoiceContext.includes("\0") ? source.agentVoiceContext.trim() : defaultSettings.agentVoiceContext,
+    interruptSpeechOnUserMessage: typeof source.interruptSpeechOnUserMessage === "boolean" ? source.interruptSpeechOnUserMessage : defaultSettings.interruptSpeechOnUserMessage,
+    sendingMode: source.sendingMode === "automatic" ? "queue" : ["manual", "queue", "steer"].includes(source.sendingMode) ? source.sendingMode : defaultSettings.sendingMode,
+    autoSendDelaySeconds: Number.isInteger(source.autoSendDelaySeconds) && source.autoSendDelaySeconds >= 2 && source.autoSendDelaySeconds <= 10 ? source.autoSendDelaySeconds : defaultSettings.autoSendDelaySeconds,
+    assistantSpeechDelaySeconds: Number.isInteger(source.assistantSpeechDelaySeconds) && source.assistantSpeechDelaySeconds >= 1 && source.assistantSpeechDelaySeconds <= 10 ? source.assistantSpeechDelaySeconds : defaultSettings.assistantSpeechDelaySeconds,
+    mode: ["speaker", "headphones"].includes(source.mode) ? source.mode : defaultSettings.mode,
+    lang: typeof source.lang === "string" && /^[a-z]{2,3}(?:-[A-Za-z0-9]{2,8})*$/.test(source.lang) ? source.lang : defaultSettings.lang,
+    recognitionLang: source.recognitionLang === "auto" || typeof source.recognitionLang === "string" && /^[a-z]{2,3}(?:-[A-Za-z0-9]{2,8})*$/.test(source.recognitionLang) ? source.recognitionLang : typeof source.lang === "string" && /^[a-z]{2,3}(?:-[A-Za-z0-9]{2,8})*$/.test(source.lang) ? source.lang : defaultSettings.recognitionLang,
+    voice: typeof source.voice === "string" && source.voice.length <= 200 && !source.voice.includes("\0") ? source.voice : "",
+    inputDeviceId: typeof source.inputDeviceId === "string" && source.inputDeviceId.length <= 500 && !source.inputDeviceId.includes("\0") ? source.inputDeviceId : "",
+    outputDeviceId: typeof source.outputDeviceId === "string" && source.outputDeviceId.length <= 500 && !source.outputDeviceId.includes("\0") ? source.outputDeviceId : "",
+    recognitionFilterEnabled: typeof source.recognitionFilterEnabled === "boolean" ? source.recognitionFilterEnabled : defaultSettings.recognitionFilterEnabled,
+    recognitionMinimumWords: Number.isInteger(source.recognitionMinimumWords) && source.recognitionMinimumWords >= 1 && source.recognitionMinimumWords <= 20 ? source.recognitionMinimumWords : defaultSettings.recognitionMinimumWords,
+    voiceCommandsEnabled: typeof source.voiceCommandsEnabled === "boolean" ? source.voiceCommandsEnabled : defaultSettings.voiceCommandsEnabled,
+    voiceCommandSend: normalizeCommandPhrases(
+      source.voiceCommandSend,
+      defaultSettings.voiceCommandSend
+    ),
+    voiceCommandQueue: normalizeCommandPhrases(
+      source.voiceCommandQueue,
+      defaultSettings.voiceCommandQueue
+    ),
+    voiceCommandEnd: normalizeCommandPhrases(
+      source.voiceCommandEnd,
+      defaultSettings.voiceCommandEnd
+    ),
+    voiceCommandMute: normalizeCommandPhrases(
+      source.voiceCommandMute,
+      defaultSettings.voiceCommandMute
+    ),
+    voiceCommandResume: normalizeCommandPhrases(
+      source.voiceCommandResume,
+      defaultSettings.voiceCommandResume
+    ),
+    voiceCommandStopSpeaking: normalizeCommandPhrases(
+      source.voiceCommandStopSpeaking,
+      defaultSettings.voiceCommandStopSpeaking
+    ),
+    voiceCommandClear: normalizeCommandPhrases(
+      source.voiceCommandClear,
+      defaultSettings.voiceCommandClear
+    ),
+    outputCodeFilterEnabled: typeof source.outputCodeFilterEnabled === "boolean" ? source.outputCodeFilterEnabled : defaultSettings.outputCodeFilterEnabled,
+    outputCodeMaxLines: Number.isInteger(source.outputCodeMaxLines) && source.outputCodeMaxLines >= 0 && source.outputCodeMaxLines <= 100 ? source.outputCodeMaxLines : defaultSettings.outputCodeMaxLines,
+    outputCodeNotice: typeof source.outputCodeNotice === "string" && source.outputCodeNotice.trim() && source.outputCodeNotice.length <= 300 && !source.outputCodeNotice.includes("\0") ? source.outputCodeNotice.trim() : defaultSettings.outputCodeNotice,
+    rate: Number.isFinite(source.rate) && source.rate >= 0.1 && source.rate <= 3 ? source.rate : 1,
+    segmentGapMs: Number.isFinite(source.segmentGapMs) && source.segmentGapMs >= 0 && source.segmentGapMs <= 2e3 ? Math.round(source.segmentGapMs) : 400
+  };
+}
 
 // src/modules/recognition/engines/whisper/whisperRecognitionHost.ts
 import { readFile, mkdir, writeFile, rename, rm } from "node:fs/promises";
@@ -304,7 +372,7 @@ var QwenHttpHost = class {
     baseUrl,
     timeoutMs = 3e5,
     fetchImpl = globalThis.fetch,
-    maxBytes = 2e6,
+    maxBytes = 5e7,
     maxSpeechBytes = 5e7,
     store
   } = {}) {
@@ -784,6 +852,89 @@ var SayEngine = class {
   }
 };
 
+// src/modules/settings/models/settingsHost.ts
+import { readFile as readFile3, mkdir as mkdir3, writeFile as writeFile3, rename as rename3, rm as rm3 } from "node:fs/promises";
+import { homedir as homedir3 } from "node:os";
+import { dirname as dirname3, join as join5 } from "node:path";
+import { randomUUID as randomUUID3 } from "node:crypto";
+function createSettingsStore(path = join5(homedir3(), ".dsh", "dsh-live-voice.settings.json")) {
+  let queue = Promise.resolve();
+  async function load() {
+    try {
+      return normalizeSettings(JSON.parse(await readFile3(path, "utf8")));
+    } catch (error) {
+      if (error.code === "ENOENT") return normalizeSettings({});
+      throw error;
+    }
+  }
+  return {
+    load: () => queue.then(load),
+    save(patch) {
+      const operation = queue.then(async () => {
+        const settings = normalizeSettings({ ...await load(), ...patch });
+        await mkdir3(dirname3(path), { recursive: true, mode: 448 });
+        const temporary = path + "." + randomUUID3() + ".tmp";
+        try {
+          await writeFile3(temporary, JSON.stringify(settings, null, 2) + "\n", {
+            mode: 384,
+            flag: "wx"
+          });
+          await rename3(temporary, path);
+        } finally {
+          await rm3(temporary, { force: true });
+        }
+        return settings;
+      });
+      queue = operation.then(
+        () => {
+        },
+        () => {
+        }
+      );
+      return operation;
+    }
+  };
+}
+
+// src/app/server/registerRoutes.ts
+var API_ROOT = "/api/dsh-live-voice";
+function registerSettingsRoute(ctx, store) {
+  const dispose = ctx.connection.fetch.register({
+    path: API_ROOT + "/settings",
+    methods: ["GET", "PUT"],
+    requestBody: "buffered",
+    fetch: async (request) => {
+      const headers = { "cache-control": "no-store" };
+      let patch;
+      if (request.method === "PUT") {
+        try {
+          const text = await request.text();
+          if (text.length > 64e3) throw new Error("too-large");
+          patch = JSON.parse(text);
+          if (!patch || typeof patch !== "object" || Array.isArray(patch))
+            throw new Error("invalid");
+        } catch {
+          return Response.json(
+            { ok: false, error: { code: "invalid-settings" } },
+            { status: 400, headers }
+          );
+        }
+      }
+      try {
+        const value = request.method === "GET" ? await store.load() : await store.save(patch);
+        return Response.json({ ok: true, value }, { headers });
+      } catch {
+        return Response.json(
+          { ok: false, error: { code: "settings-unavailable" } },
+          { status: 500, headers }
+        );
+      }
+    }
+  });
+  ctx.effect(() => () => dispose(), "dsh-live-voice: remove settings route");
+}
+var APP_VOICE_CONTEXT_PATH = API_ROOT + "/voice-context";
+
 // src/app/server/apply.ts
 var name = "dsh-live-voice";
 var inject = ["connection", "systemPrompt"];
@@ -880,6 +1031,7 @@ function createSayHost({ engine = new SayEngine() } = {}) {
   return { handle, dispose, engine };
 }
 function apply(ctx, {
+  settingsStore = createSettingsStore(),
   whisperStore = createWhisperConfigStore(),
   whisperFetch = globalThis.fetch,
   qwenStore = createQwenConfigStore(),
@@ -888,6 +1040,7 @@ function apply(ctx, {
   createSayEngine = () => new SayEngine(),
   encodeHostSpeech = wavToM4aAac
 } = {}) {
+  registerSettingsRoute(ctx, settingsStore);
   ctx.systemPrompt.variable(
     "live_voice_context",
     (assemblyContext) => voiceContextStore.get(String(assemblyContext.agent?.sessionId || ""))

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "dsh-live-voice",
-  "version": "0.3.0",
-  "dshTestedVersion": "0.1.6-alpha.2",
+  "version": "0.3.2",
+  "dshTestedVersion": "0.2.0-rc.2",
   "type": "module",
   "scripts": {
     "typecheck": "tsc --noEmit",

package/src/app/AGENTS.md

@@ -0,0 +1,7 @@
+# Application boundary
+
+Read [ARCHITECTURE.md](ARCHITECTURE.md) before changing composition roots, routes, slots, or language registration. Read the repository [AGENTS.md](../../AGENTS.md) for product, validation, and authorization rules.
+
+- Inspect the actual DSH extension API before changing an integration point. Preserve established slot names, IDs, ordering, injected services, authenticated same-origin routes, and browser/host separation.
+- Keep UI copy in the typed catalogs under `client/i18n/`; update every supported language in alphabetical key order. Use DSH `ctx.locale`, not an independent preference. UI locale is independent of recognition/synthesis languages and configured voice-command phrases.
+- Treat `package.json` as the source of truth for plugin and tested-DSH versions. Never hard-code these in the client.

package/src/app/ARCHITECTURE.md

@@ -0,0 +1,10 @@
+# Application architecture
+
+`src/app` is the composition boundary between DSH and the voice-domain modules. Client and server entry points remain separate; host-only Node dependencies must never enter the browser bundle.
+
+- `client/apply.tsx` injects DSH client services, creates session controllers and voice ownership, wires engines and settings, and disposes resources. `client/slotDefinitions.ts` and `client/registerSlots.tsx` own the established slot IDs, order, and registration. Preserve those contracts when changing placement.
+- `server/apply.ts` composes host engines and providers. `server/registerRoutes.ts` registers authenticated same-origin routes; do not bypass their validation or expose host access directly to browser code.
+- `client/i18n` owns the only UI translation tree: typed, alphabetized catalogs for every supported locale, runtime selection synchronized with `ctx.locale`, DSH registration, and the boundary wrapping slot components. UI language must not change STT/TTS language or user-defined command phrases.
+- Styles are assembled from `src/styles/index.ts`; design-system primitives live under `src/shared/design-system` rather than here.
+
+Read [the repository overview](../../docs/ARCHITECTURE.md) for dependency direction and stable contracts, and the relevant module architecture before editing feature behavior. Read [application agent guidance](AGENTS.md) for operational constraints.

package/src/app/client/apply.tsx

@@ -1,12 +1,7 @@
 // @ts-nocheck
 import React from 'react';
 import { createPortal } from 'react-dom';
-import {
-  MicrophoneMeter,
-  VoiceCoordinator,
-  VoiceOwnership,
-  normalizeSettings,
-} from '../../modules/core/index.js';
+import { MicrophoneMeter, VoiceCoordinator, VoiceOwnership } from '../../modules/core/index.js';
 import {
   BrowserRecognitionEngine,
   QwenHttpRecognitionEngine,
@@ -26,6 +21,7 @@ import {
 import { registerLiveVoiceLocales } from './i18n/index.js';
 import { registerConversationSlots, registerSettingsSlot } from './registerSlots.js';
 import { styles } from '../../styles/index.js';
+import { createSettingsClient } from '../../modules/settings/models/settingsStorage.js';
 import {
   assistantMessages,
   addressedTurn,
@@ -39,6 +35,8 @@ export function apply(ctx) {
   const SettingsPanel = createLiveVoiceSettings(t, ctx.locale);
   const controllers = new Map();
   const retiring = new Set();
+  const preferences = createSettingsClient();
+  const settingsControllers = new Set();
   let disposed = false;
   // Voice conversation mode belongs to the plugin, not to a mounted composer.
   // A route change may replace every conversation slot, but the next committed
@@ -73,6 +71,7 @@ export function apply(ctx) {
     'recognitionProcessLocally',
     'recognitionAutoInstall',
     'voiceDetectionPreset',
+    'voiceDetectionCustomSilenceMs',
     'recognitionMaxUtteranceSeconds',
   ]);
   const changesRecognition = (next) =>
@@ -82,12 +81,14 @@ export function apply(ctx) {
       ? new QwenHttpRecognitionEngine({
           meter,
           voiceDetectionPreset: settings.voiceDetectionPreset,
+          voiceDetectionCustomSilenceMs: settings.voiceDetectionCustomSilenceMs,
           maxUtteranceSeconds: settings.recognitionMaxUtteranceSeconds,
         })
       : settings.recognitionEngine === 'whisper-http'
         ? new WhisperHttpRecognitionEngine({
             meter,
             voiceDetectionPreset: settings.voiceDetectionPreset,
+            voiceDetectionCustomSilenceMs: settings.voiceDetectionCustomSilenceMs,
             maxUtteranceSeconds: settings.recognitionMaxUtteranceSeconds,
           })
         : new BrowserRecognitionEngine({
@@ -99,6 +100,38 @@ export function apply(ctx) {
       if (!disposed && !controller.disposed)
         controller.patch({ error: error?.message || String(error) });
     });
+  async function savePreferences(controller, next) {
+    try {
+      return await preferences.save(next);
+    } catch {
+      const error = new Error(t('dsh-live-voice.settings.persistence.saveError'));
+      if (!disposed && !controller.disposed) controller.patch({ error: error.message });
+      throw error;
+    }
+  }
+  const unsubscribePreferences = preferences.subscribe((settings) => {
+    if (disposed) return;
+    for (const entry of controllers.values()) {
+      if (entry.closed) continue;
+      const previous = entry.controller.getSnapshot().settings;
+      if ([...recognitionSettingKeys].some((key) => previous[key] !== settings[key]))
+        entry.controller.replaceRecognition(recognitionFor(settings, entry.controller.meter));
+      entry.applySettings(settings);
+    }
+    for (const c of settingsControllers) {
+      if (c.disposed) continue;
+      const previous = c.getSnapshot().settings;
+      if ([...recognitionSettingKeys].some((key) => previous[key] !== settings[key]))
+        c.replaceRecognition(recognitionFor(settings, c.meter));
+      c.applySavedSettings(settings);
+      for (const engine of Object.values(c.engines)) engine.lang = settings.lang;
+      run(c, c.refreshCapabilities());
+    }
+  });
+  ctx.effect(
+    () => () => unsubscribePreferences(),
+    'dsh-live-voice: remove preferences subscription',
+  );
   function retire(entry) {
     if (entry.closed) return;
     entry.closed = true;
@@ -128,6 +161,10 @@ export function apply(ctx) {
       key,
       draft: '',
       pendingDraft: undefined,
+      publishedDraft: undefined,
+      publishedRevision: undefined,
+      pendingRevision: undefined,
+      staleDrafts: new Set(),
       composers: new Map(),
       refs: 0,
       buttons: 0,
@@ -135,14 +172,7 @@ export function apply(ctx) {
       closed: false,
       lastVoiceContextBody: null,
     };
-    let settings = {};
-    try {
-      settings = normalizeSettings(
-        JSON.parse(localStorage.getItem('dsh-live-voice.settings') || '{}'),
-      );
-    } catch {
-      settings = normalizeSettings(null);
-    }
+    const settings = preferences.getSnapshot();
     const engineBrowser = new BrowserSpeakingEngine({ lang: settings.lang || 'pt-BR' });
     const engineSay = new HostAudioSpeakingEngine({
       endpoint: '/api/dsh-live-voice/say/speech',
@@ -202,12 +232,14 @@ export function apply(ctx) {
           if (disposed || entry.closed) return;
           const owner = [...entry.composers.values()].at(-1);
           if (!owner) return;
-          // `setDraft()` updates Lexical synchronously but the subscribed InputState
-          // can publish on a subsequent React commit. Preserve the optimistic value
-          // until that exact publication arrives; otherwise another slot render can
-          // make a later recognition result start from stale text.
+          // Preserve unacknowledged voice writes across unchanged slot renders,
+          // but let new editor publications (including manual edits) supersede them.
+          if (entry.pendingDraft === undefined) entry.staleDrafts.clear();
+          entry.staleDrafts.add(entry.draft);
+          if (entry.publishedDraft !== undefined) entry.staleDrafts.add(entry.publishedDraft);
           entry.draft = text;
           entry.pendingDraft = text;
+          entry.pendingRevision = entry.publishedRevision;
           owner.actions.setDraft(text);
         },
       },
@@ -286,13 +318,7 @@ export function apply(ctx) {
     controller.updateSettings = (next) => {
       if (disposed || entry.closed) return;
       entry.applySettings(next);
-      const settings = controller.getSnapshot().settings;
-      try {
-        localStorage.setItem('dsh-live-voice.settings', JSON.stringify(settings));
-      } catch {}
-      for (const other of controllers.values()) {
-        if (other !== entry && !other.closed) other.applySettings(settings);
-      }
+      return savePreferences(controller, next);
     };
     // Cancel only this entry's queued acquisition. Global cancellation here would
     // invalidate a newer session while its predecessor is being unmounted.
@@ -322,7 +348,12 @@ export function apply(ctx) {
         return ownership.run(
           controller,
           [...controllers.values()].map((other) => other.controller).concat([...retiring]),
-          () => {
+          async () => {
+            try {
+              await preferences.ready;
+            } catch {
+              throw new Error(t('dsh-live-voice.settings.persistence.loadError'));
+            }
             if (disposed || entry.closed || !entry.refs || request !== entry.request) return;
             if (method !== 'speak' && !entry.composers.size) return;
             const result = original(...args);
@@ -333,6 +364,10 @@ export function apply(ctx) {
       };
     }
     controllers.set(key, entry);
+    void preferences.ready.catch(() => {
+      if (!entry.closed)
+        controller.patch({ error: t('dsh-live-voice.settings.persistence.loadError') });
+    });
     run(controller, controller.refreshCapabilities());
     return entry;
   }
@@ -409,10 +444,22 @@ export function apply(ctx) {
       const published = typeof input.draft === 'string' ? input.draft : '';
       // A voice write is optimistic until React commits the editor publication.
       // Unrelated slot renders must not restore the previous draft in between.
+      entry.publishedDraft = published;
+      entry.publishedRevision = input.draftRev;
+      const newerRevision =
+        Number.isInteger(input.draftRev) &&
+        Number.isInteger(entry.pendingRevision) &&
+        input.draftRev > entry.pendingRevision;
       if (entry.pendingDraft === undefined) entry.draft = published;
-      else if (published === entry.pendingDraft) {
+      else if (
+        published === entry.pendingDraft ||
+        newerRevision ||
+        !entry.staleDrafts.has(published)
+      ) {
+        // A new manual edit is authoritative even if the exact voice echo was skipped.
         entry.draft = published;
         entry.pendingDraft = undefined;
+        entry.staleDrafts.clear();
       }
       entry.controller.composerChanged(entry.draft);
     });
@@ -425,14 +472,7 @@ export function apply(ctx) {
   function Settings() {
     const [controller, setController] = React.useState(null);
     React.useEffect(() => {
-      let settings;
-      try {
-        settings = normalizeSettings(
-          JSON.parse(localStorage.getItem('dsh-live-voice.settings') || '{}'),
-        );
-      } catch {
-        settings = normalizeSettings(null);
-      }
+      const settings = preferences.getSnapshot();
       const browser = new BrowserSpeakingEngine({ lang: settings.lang });
       const qwen = new QwenHttpSpeakingEngine({ lang: settings.lang });
       const meter = new MicrophoneMeter();
@@ -458,9 +498,21 @@ export function apply(ctx) {
         ownership.run(
           c,
           [...controllers.values()].map((entry) => entry.controller).concat([...retiring]),
-          () => speak(...args),
+          async () => {
+            try {
+              await preferences.ready;
+            } catch {
+              throw new Error(t('dsh-live-voice.settings.persistence.loadError'));
+            }
+            if (!disposed && !c.disposed) return speak(...args);
+          },
         );
       const update = c.updateSettings.bind(c);
+      c.applySavedSettings = update;
+      settingsControllers.add(c);
+      void preferences.ready.catch(() =>
+        c.patch({ error: t('dsh-live-voice.settings.persistence.loadError') }),
+      );
       let settingsRevision = 0;
       c.updateSettings = async (next) => {
         const revision = ++settingsRevision;
@@ -469,10 +521,11 @@ export function apply(ctx) {
         qwen.lang = c.getSnapshot().settings.lang;
         const settings = c.getSnapshot().settings;
         if (changesRecognition(next)) c.replaceRecognition(recognitionFor(settings, c.meter));
-        localStorage.setItem('dsh-live-voice.settings', JSON.stringify(settings));
+        const saved = savePreferences(c, next);
         run(c, c.refreshCapabilities());
         const active = [...controllers.values()];
-        await Promise.allSettled([
+        const [savedSettings] = await Promise.all([
+          saved,
           c.endConversation(),
           ...active.map((entry) => entry.controller.endConversation()),
         ]);
@@ -480,8 +533,10 @@ export function apply(ctx) {
         for (const entry of controllers.values()) {
           if (entry.closed) continue;
           if (changesRecognition(next))
-            entry.controller.replaceRecognition(recognitionFor(settings, entry.controller.meter));
-          entry.applySettings(settings);
+            entry.controller.replaceRecognition(
+              recognitionFor(savedSettings, entry.controller.meter),
+            );
+          entry.applySettings(savedSettings);
           run(entry.controller, entry.controller.refreshCapabilities());
         }
       };
@@ -491,6 +546,7 @@ export function apply(ctx) {
       refresh();
       return () => {
         document.removeEventListener('dsh-live-voice:capabilitieschanged', refresh);
+        settingsControllers.delete(c);
         void c.dispose();
       };
     }, []);

package/src/app/client/i18n/catalogs/base.ts

@@ -88,6 +88,8 @@ export interface LiveVoiceTranslation extends Translation {
   'dsh-live-voice.recognition.planned.vote': string;
   'dsh-live-voice.recognition.planned.voxtral': string;
   'dsh-live-voice.recognition.planned.webGpu': string;
+  'dsh-live-voice.recognition.presets.custom.description': string;
+  'dsh-live-voice.recognition.presets.custom.label': string;
   'dsh-live-voice.recognition.presets.long.description': string;
   'dsh-live-voice.recognition.presets.long.label': string;
   'dsh-live-voice.recognition.presets.natural.description': string;
@@ -100,6 +102,8 @@ export interface LiveVoiceTranslation extends Translation {
   'dsh-live-voice.recognition.qwen.endpointHelp': string;
   'dsh-live-voice.recognition.qwen.hostHelp': string;
   'dsh-live-voice.recognition.qwen.label': string;
+  'dsh-live-voice.recognition.silenceDetection.customHelp': string;
+  'dsh-live-voice.recognition.silenceDetection.customLabel': string;
   'dsh-live-voice.recognition.silenceDetection.duration': string;
   'dsh-live-voice.recognition.silenceDetection.help': string;
   'dsh-live-voice.recognition.silenceDetection.label': string;
@@ -136,6 +140,9 @@ export interface LiveVoiceTranslation extends Translation {
   'dsh-live-voice.settings.delivery.toggle': string;
   'dsh-live-voice.settings.engine.refresh': string;
   'dsh-live-voice.settings.filters.title': string;
+  'dsh-live-voice.settings.general.title': string;
+  'dsh-live-voice.settings.persistence.loadError': string;
+  'dsh-live-voice.settings.persistence.saveError': string;
   'dsh-live-voice.settings.tabs.conversation': string;
   'dsh-live-voice.settings.tabs.recognition': string;
   'dsh-live-voice.settings.tabs.speak': string;

package/src/app/client/i18n/catalogs/en.ts

@@ -101,6 +101,8 @@ const en: LiveVoiceTranslation = {
   'dsh-live-voice.recognition.planned.vote': 'Coming Soon — vote on repo issues',
   'dsh-live-voice.recognition.planned.voxtral': 'Voxtral Realtime — Soon',
   'dsh-live-voice.recognition.planned.webGpu': 'Browser WebGPU Inference — Soon',
+  'dsh-live-voice.recognition.presets.custom.description': 'Choose your own silence duration.',
+  'dsh-live-voice.recognition.presets.custom.label': 'Custom',
   'dsh-live-voice.recognition.presets.long.description': 'Wait through longer thinking pauses.',
   'dsh-live-voice.recognition.presets.long.label': 'Long',
   'dsh-live-voice.recognition.presets.natural.description': 'Allow normal pauses between phrases.',
@@ -118,6 +120,9 @@ const en: LiveVoiceTranslation = {
   'dsh-live-voice.recognition.qwen.hostHelp':
     'Host-wide settings for the Qwen3 ASR + TTS server. Enter any HTTP or HTTPS base URL reachable from the DSH host. The browser accesses it through authenticated DSH routes.',
   'dsh-live-voice.recognition.qwen.label': 'Qwen3 ASR — HTTP API',
+  'dsh-live-voice.recognition.silenceDetection.customHelp':
+    'Enter a whole number from 100 to 10,000 ms. Saved when you leave the field. Short pauses may split speech; recognition adds its own latency.',
+  'dsh-live-voice.recognition.silenceDetection.customLabel': 'Custom pause (milliseconds)',
   'dsh-live-voice.recognition.silenceDetection.duration': 'Pause before sending: {milliseconds} ms',
   'dsh-live-voice.recognition.silenceDetection.help':
     'Controls how long a pause must last before captured speech is sent for recognition.',
@@ -162,6 +167,11 @@ const en: LiveVoiceTranslation = {
   'dsh-live-voice.settings.delivery.toggle': 'Automatic delivery mode',
   'dsh-live-voice.settings.engine.refresh': 'Refresh available engines',
   'dsh-live-voice.settings.filters.title': 'Filtering',
+  'dsh-live-voice.settings.general.title': 'General',
+  'dsh-live-voice.settings.persistence.loadError':
+    'Could not load Live Voice settings from the server. Reload to try again.',
+  'dsh-live-voice.settings.persistence.saveError':
+    'Could not save Live Voice settings on the server. Please try again.',
   'dsh-live-voice.settings.tabs.conversation': 'Conversation',
   'dsh-live-voice.settings.tabs.recognition': 'Speech recognition',
   'dsh-live-voice.settings.tabs.speak': 'Speech',

package/src/app/client/i18n/catalogs/es.ts

@@ -107,6 +107,8 @@ const es: LiveVoiceTranslation = {
     'Próximamente — vota en las incidencias del repositorio',
   'dsh-live-voice.recognition.planned.voxtral': 'Voxtral Realtime — Próximamente',
   'dsh-live-voice.recognition.planned.webGpu': 'Inferencia WebGPU en el navegador — Próximamente',
+  'dsh-live-voice.recognition.presets.custom.description': 'Elige la duración del silencio.',
+  'dsh-live-voice.recognition.presets.custom.label': 'Personalizada',
   'dsh-live-voice.recognition.presets.long.description':
     'Espera durante pausas de reflexión más largas.',
   'dsh-live-voice.recognition.presets.long.label': 'Larga',
@@ -126,6 +128,9 @@ const es: LiveVoiceTranslation = {
   'dsh-live-voice.recognition.qwen.hostHelp':
     'Configuración para todo el host del servidor Qwen3 ASR + TTS. Introduce una URL base HTTP o HTTPS accesible desde el host de DSH. El navegador accede a ella mediante las rutas autenticadas de DSH.',
   'dsh-live-voice.recognition.qwen.label': 'Qwen3 ASR — API HTTP',
+  'dsh-live-voice.recognition.silenceDetection.customHelp':
+    'Introduce un número entero de 100 a 10.000 ms. Se guarda al salir del campo. Las pausas cortas pueden dividir el habla; el reconocimiento añade su propia latencia.',
+  'dsh-live-voice.recognition.silenceDetection.customLabel': 'Pausa personalizada (milisegundos)',
   'dsh-live-voice.recognition.silenceDetection.duration':
     'Pausa antes de enviar: {milliseconds} ms',
   'dsh-live-voice.recognition.silenceDetection.help':
@@ -175,6 +180,11 @@ const es: LiveVoiceTranslation = {
   'dsh-live-voice.settings.delivery.toggle': 'Modo de envío automático',
   'dsh-live-voice.settings.engine.refresh': 'Actualizar los motores disponibles',
   'dsh-live-voice.settings.filters.title': 'Filtrado',
+  'dsh-live-voice.settings.general.title': 'General',
+  'dsh-live-voice.settings.persistence.loadError':
+    'No se pudieron cargar los ajustes de Live Voice del servidor. Recarga para volver a intentarlo.',
+  'dsh-live-voice.settings.persistence.saveError':
+    'No se pudieron guardar los ajustes de Live Voice en el servidor. Inténtalo de nuevo.',
   'dsh-live-voice.settings.tabs.conversation': 'Conversación',
   'dsh-live-voice.settings.tabs.recognition': 'Reconocimiento de voz',
   'dsh-live-voice.settings.tabs.speak': 'Síntesis de voz',

package/src/app/client/i18n/catalogs/fr.ts

@@ -109,6 +109,8 @@ const fr: LiveVoiceTranslation = {
   'dsh-live-voice.recognition.planned.voxtral': 'Voxtral Realtime — Bientôt disponible',
   'dsh-live-voice.recognition.planned.webGpu':
     'Inférence WebGPU dans le navigateur — Bientôt disponible',
+  'dsh-live-voice.recognition.presets.custom.description': 'Choisissez la durée du silence.',
+  'dsh-live-voice.recognition.presets.custom.label': 'Personnalisée',
   'dsh-live-voice.recognition.presets.long.description':
     'Attend pendant les pauses de réflexion plus longues.',
   'dsh-live-voice.recognition.presets.long.label': 'Longue',
@@ -129,6 +131,9 @@ const fr: LiveVoiceTranslation = {
   'dsh-live-voice.recognition.qwen.hostHelp':
     'Paramètres communs à tout l’hôte pour le serveur Qwen3 ASR + TTS. Saisissez une URL de base HTTP ou HTTPS accessible depuis l’hôte DSH. Le navigateur y accède via les routes authentifiées de DSH.',
   'dsh-live-voice.recognition.qwen.label': 'Qwen3 ASR — API HTTP',
+  'dsh-live-voice.recognition.silenceDetection.customHelp':
+    'Saisissez un entier de 100 à 10 000 ms. Enregistré en quittant le champ. Les pauses courtes peuvent couper la parole ; la reconnaissance ajoute sa propre latence.',
+  'dsh-live-voice.recognition.silenceDetection.customLabel': 'Pause personnalisée (millisecondes)',
   'dsh-live-voice.recognition.silenceDetection.duration': 'Pause avant l’envoi : {milliseconds} ms',
   'dsh-live-voice.recognition.silenceDetection.help':
     'Détermine la durée de pause nécessaire avant que la parole capturée soit envoyée pour reconnaissance.',
@@ -178,6 +183,11 @@ const fr: LiveVoiceTranslation = {
   'dsh-live-voice.settings.delivery.toggle': 'Mode d’envoi automatique',
   'dsh-live-voice.settings.engine.refresh': 'Actualiser les moteurs disponibles',
   'dsh-live-voice.settings.filters.title': 'Filtrage',
+  'dsh-live-voice.settings.general.title': 'Général',
+  'dsh-live-voice.settings.persistence.loadError':
+    'Impossible de charger les paramètres Live Voice depuis le serveur. Rechargez pour réessayer.',
+  'dsh-live-voice.settings.persistence.saveError':
+    'Impossible de sauvegarder les paramètres Live Voice sur le serveur. Réessayez.',
   'dsh-live-voice.settings.tabs.conversation': 'Conversation',
   'dsh-live-voice.settings.tabs.recognition': 'Reconnaissance vocale',
   'dsh-live-voice.settings.tabs.speak': 'Synthèse vocale',