dsh-live-voice 0.6.1 → 0.6.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (34) hide show
  1. package/README.md +1 -1
  2. package/lib/asr/asr.worker.js +1 -1
  3. package/lib/asr/manifest.json +1 -1
  4. package/lib/client.js +400 -66
  5. package/lib/server.js +47 -7
  6. package/lib/vad/manifest.json +2 -2
  7. package/lib/vad/vad.worker.js +1 -1
  8. package/package.json +2 -2
  9. package/src/app/client/apply.tsx +14 -0
  10. package/src/app/client/i18n/catalogs/base.ts +4 -1
  11. package/src/app/client/i18n/catalogs/en.ts +6 -1
  12. package/src/app/client/i18n/catalogs/es.ts +6 -1
  13. package/src/app/client/i18n/catalogs/fr.ts +6 -1
  14. package/src/app/client/i18n/catalogs/hi.ts +6 -1
  15. package/src/app/client/i18n/catalogs/pt-BR.ts +6 -1
  16. package/src/app/client/i18n/catalogs/zh.ts +6 -1
  17. package/src/app/server/apply.ts +19 -0
  18. package/src/modules/core/coordinator.ts +45 -4
  19. package/src/modules/core/qwen/QwenHttpHost.ts +5 -1
  20. package/src/modules/core/settings.ts +17 -7
  21. package/src/modules/recognition/ARCHITECTURE.md +3 -1
  22. package/src/modules/recognition/engines/whisper/WhisperRecognitionEngine.ts +2 -0
  23. package/src/modules/recognition/engines/whisper/whisperRecognitionHost.ts +4 -1
  24. package/src/modules/recognition/local/BrowserLocalRecognitionEngine.ts +4 -1
  25. package/src/modules/recognition/transcriptionContext.ts +28 -0
  26. package/src/modules/recording/ARCHITECTURE.md +7 -1
  27. package/src/modules/recording/PinDetector.ts +6 -0
  28. package/src/modules/recording/RecordingRecognition.ts +155 -38
  29. package/src/modules/recording/codec.ts +4 -0
  30. package/src/modules/recording/repository.ts +137 -5
  31. package/src/modules/recording/storage.ts +21 -1
  32. package/src/modules/recording/types.ts +19 -0
  33. package/src/modules/settings/sections/conversation/ConversationDelaySettings.tsx +9 -7
  34. package/src/modules/settings/sections/recognition/RecognitionEngineSettings.tsx +17 -9
package/lib/client.js CHANGED
@@ -152,7 +152,7 @@ var SessionSpeechMemory = class {
152
152
  };
153
153
 
154
154
  // src/modules/recognition/vad/assets.ts
155
- var sileroVadAssetBase = () => "/api/dsh-live-voice/vad/" + encodeURIComponent(true ? "0.6.1" : "test") + "/";
155
+ var sileroVadAssetBase = () => "/api/dsh-live-voice/vad/" + encodeURIComponent(true ? "0.6.2" : "test") + "/";
156
156
  async function sileroVadAvailable(fetchImpl = globalThis.fetch) {
157
157
  try {
158
158
  const response = await fetchImpl(sileroVadAssetBase() + "availability", {
@@ -615,6 +615,7 @@ var RecordingRepositoryError = class extends Error {
615
615
  }
616
616
  code;
617
617
  };
618
+ var audioPageBytes = (page) => (page.decoderConfig?.description?.byteLength ?? 0) + page.packets.reduce((total, packet) => total + packet.data.byteLength, 0);
618
619
  var globals = () => globalThis;
619
620
  function unsupported(cause) {
620
621
  return new RecordingRepositoryError(
@@ -919,6 +920,19 @@ async function openRecordingStore() {
919
920
  }
920
921
  };
921
922
  }),
923
+ audioBytes: (id) => transact("readonly", (tx, result) => {
924
+ let bytes = 0;
925
+ const request = tx.objectStore("pages").index("recordingId").openCursor(IDBKeyRange.only(id));
926
+ request.onsuccess = () => {
927
+ const cursor = request.result;
928
+ if (!cursor) {
929
+ result(bytes);
930
+ return;
931
+ }
932
+ bytes += audioPageBytes(cursor.value);
933
+ cursor.continue();
934
+ };
935
+ }),
922
936
  nextPage: (id, start, end) => transact("readonly", (tx, result) => {
923
937
  if (end <= start) {
924
938
  result(null);
@@ -956,6 +970,7 @@ function createMemoryStore(maxBytes = 32 * 1024 * 1024) {
956
970
  let used = 0;
957
971
  const size2 = (page) => 256 + (page.decoderConfig?.description?.byteLength ?? 0) + page.packets.reduce((total, packet) => total + packet.data.byteLength + 64, 0);
958
972
  return {
973
+ audioBytes: async (id) => (pages.get(id) ?? []).reduce((sum, page) => sum + audioPageBytes(page), 0),
959
974
  list: async () => structuredClone([...sessions.values()]),
960
975
  get: async (id) => structuredClone(sessions.get(id) ?? null),
961
976
  create: async (session) => {
@@ -1204,6 +1219,17 @@ function createRecordingRepository() {
1204
1219
  }
1205
1220
  return locate(id);
1206
1221
  }
1222
+ const measured = /* @__PURE__ */ new Map();
1223
+ async function withAudioBytes(session) {
1224
+ if (session.audioBytes !== void 0) return session;
1225
+ let cached = measured.get(session.id);
1226
+ if (!cached || cached.samples !== session.savedSamples) {
1227
+ const { store } = await locate(session.id);
1228
+ cached = { samples: session.savedSamples, bytes: await store.audioBytes(session.id) };
1229
+ measured.set(session.id, cached);
1230
+ }
1231
+ return { ...session, audioBytes: cached.bytes };
1232
+ }
1207
1233
  const repository = {
1208
1234
  version: 1,
1209
1235
  segments: recordingSegments,
@@ -1214,14 +1240,21 @@ function createRecordingRepository() {
1214
1240
  },
1215
1241
  async list() {
1216
1242
  check2();
1217
- return (await all()).map(
1218
- (session) => writers.get(session.id)?.writer.session ?? cloneSession(session)
1243
+ const sessions = await all();
1244
+ const ids = new Set(sessions.map((session) => session.id));
1245
+ for (const id of measured.keys()) if (!ids.has(id)) measured.delete(id);
1246
+ return Promise.all(
1247
+ sessions.map(
1248
+ (session) => withAudioBytes(writers.get(session.id)?.writer.session ?? cloneSession(session))
1249
+ )
1219
1250
  );
1220
1251
  },
1221
1252
  async get(id) {
1222
1253
  check2();
1223
1254
  try {
1224
- return writers.get(id)?.writer.session ?? cloneSession((await locate(id)).session);
1255
+ return await withAudioBytes(
1256
+ writers.get(id)?.writer.session ?? cloneSession((await locate(id)).session)
1257
+ );
1225
1258
  } catch (error) {
1226
1259
  if (error instanceof RecordingRepositoryError && error.code === "not-found") return null;
1227
1260
  throw error;
@@ -1326,6 +1359,66 @@ function createRecordingRepository() {
1326
1359
  emit();
1327
1360
  });
1328
1361
  },
1362
+ async setName(id, value) {
1363
+ check2();
1364
+ if (typeof value !== "string" || value.trim().length > 120 || /[\u0000-\u001f\u007f-\u009f\u2028\u2029]/u.test(value))
1365
+ throw new RecordingRepositoryError(
1366
+ "invalid-range",
1367
+ "Recording names must be a single line of at most 120 characters."
1368
+ );
1369
+ const name = value.trim();
1370
+ const active = writers.get(id);
1371
+ if (active) return active.setName(name);
1372
+ await serial(async () => {
1373
+ const { store } = await locate(id);
1374
+ const release = store === memory ? () => {
1375
+ } : await holdLock(id);
1376
+ try {
1377
+ await store.update(id, (current) => {
1378
+ if (!current.starred)
1379
+ throw new RecordingRepositoryError(
1380
+ "invalid-range",
1381
+ "Star a recording before naming it."
1382
+ );
1383
+ return { ...current, name };
1384
+ });
1385
+ } finally {
1386
+ release();
1387
+ }
1388
+ emit();
1389
+ });
1390
+ },
1391
+ timeline() {
1392
+ const targets = [...writers.entries()];
1393
+ return async (kind) => {
1394
+ const occurredAt = now();
1395
+ await Promise.all(
1396
+ targets.map(async ([id, active]) => {
1397
+ const value = {
1398
+ id: crypto.randomUUID(),
1399
+ kind,
1400
+ occurredAt,
1401
+ atSample: active.writer.session.capturedSamples
1402
+ };
1403
+ if (writers.get(id) === active) return active.event(value);
1404
+ await serial(async () => {
1405
+ const { store } = await locate(id);
1406
+ const release = store === memory ? () => {
1407
+ } : await holdLock(id);
1408
+ try {
1409
+ await store.update(id, (current) => ({
1410
+ ...current,
1411
+ events: [...current.events ?? [], value]
1412
+ }));
1413
+ } finally {
1414
+ release();
1415
+ }
1416
+ emit();
1417
+ });
1418
+ })
1419
+ );
1420
+ };
1421
+ },
1329
1422
  async begin(input) {
1330
1423
  check2();
1331
1424
  return serial(async () => {
@@ -1363,12 +1456,16 @@ function createRecordingRepository() {
1363
1456
  state: "recording",
1364
1457
  capturedSamples: 0,
1365
1458
  savedSamples: 0,
1459
+ audioBytes: 0,
1366
1460
  persistent,
1367
1461
  codec: "opus",
1368
1462
  storage: policy.enabled ? "indexeddb" : "memory",
1369
1463
  configuration: input.configuration,
1370
1464
  pins: [],
1371
1465
  attempts: [],
1466
+ events: [
1467
+ { id: crypto.randomUUID(), kind: "capture-start", occurredAt: now(), atSample: 0 }
1468
+ ],
1372
1469
  keepPrevious: Math.max(1, Math.floor(policy.keepPrevious))
1373
1470
  };
1374
1471
  try {
@@ -1422,7 +1519,14 @@ function createRecordingRepository() {
1422
1519
  sampleRate,
1423
1520
  ...encoded
1424
1521
  };
1425
- await update((current) => ({ ...current, savedSamples: page.startSample + count }), page);
1522
+ await update(
1523
+ (current) => ({
1524
+ ...current,
1525
+ savedSamples: page.startSample + count,
1526
+ audioBytes: (current.audioBytes ?? 0) + audioPageBytes(page)
1527
+ }),
1528
+ page
1529
+ );
1426
1530
  pendingCount = 0;
1427
1531
  index++;
1428
1532
  };
@@ -1553,7 +1657,16 @@ function createRecordingRepository() {
1553
1657
  ...current,
1554
1658
  state: fatal ? "failed" : state,
1555
1659
  error: fatal?.code ?? error,
1556
- endedAt: now()
1660
+ endedAt: now(),
1661
+ events: [
1662
+ ...current.events ?? [],
1663
+ {
1664
+ id: crypto.randomUUID(),
1665
+ kind: "capture-end",
1666
+ occurredAt: now(),
1667
+ atSample: current.capturedSamples
1668
+ }
1669
+ ]
1557
1670
  }));
1558
1671
  } finally {
1559
1672
  pending = new Float32Array(0);
@@ -1572,6 +1685,25 @@ function createRecordingRepository() {
1572
1685
  setStarred: (value) => enqueue(async () => {
1573
1686
  await update((current) => ({ ...current, starred: value }));
1574
1687
  }),
1688
+ setName: (name) => enqueue(async () => {
1689
+ await update((current) => {
1690
+ if (!current.starred)
1691
+ throw new RecordingRepositoryError(
1692
+ "invalid-range",
1693
+ "Star a recording before naming it."
1694
+ );
1695
+ return { ...current, name };
1696
+ });
1697
+ }),
1698
+ event: (value) => enqueue(async () => {
1699
+ await update((current) => ({
1700
+ ...current,
1701
+ events: [
1702
+ ...current.events ?? [],
1703
+ { ...value, atSample: current.capturedSamples }
1704
+ ]
1705
+ }));
1706
+ }),
1575
1707
  idle: async () => {
1576
1708
  await tail;
1577
1709
  }
@@ -1807,6 +1939,9 @@ var en = {
1807
1939
  "dsh-live-voice.recognition.commands.send": "Send to running agent",
1808
1940
  "dsh-live-voice.recognition.commands.stopSpeech": "Stop assistant speech",
1809
1941
  "dsh-live-voice.recognition.commands.title": "Voice commands",
1942
+ "dsh-live-voice.recognition.context.hint": "Optional background and vocabulary sent as prompt to the selected HTTP speech service, together with each audio segment. Saved on blur. Support depends on the server; this does not restrict languages or guarantee instruction following. Leave empty to send no context.",
1943
+ "dsh-live-voice.recognition.context.label": "Transcription context",
1944
+ "dsh-live-voice.recognition.context.placeholder": "Native Brazilian Portuguese speaker who frequently uses technical terms in English.",
1810
1945
  "dsh-live-voice.recognition.dictation.cancel": "Cancel dictation",
1811
1946
  "dsh-live-voice.recognition.engine.label": "Recognition engine",
1812
1947
  "dsh-live-voice.recognition.headphoneMode.help": "The microphone stays open during playback. Recognized non-echo speech can pause playback after the configured delay; it resumes after microphone and shared meeting audio stay silent for that delay. Manual Pause requires Resume.",
@@ -2079,7 +2214,7 @@ var en = {
2079
2214
  "dsh-live-voice.speak.rate.label": "Speech rate",
2080
2215
  "dsh-live-voice.speak.responseDelay.help": "After you stop speaking, automatic assistant playback waits for this much continuous silence. Speaking again restarts the wait.",
2081
2216
  "dsh-live-voice.speak.responseDelay.label": "Assistant response delay",
2082
- "dsh-live-voice.speak.responseDelay.none": "No delay",
2217
+ "dsh-live-voice.speak.responseDelay.oneAndHalf": "1.5 seconds",
2083
2218
  "dsh-live-voice.speak.segmentGap.help": "Wait this many milliseconds between consecutive spoken segments. 400 ms is the default.",
2084
2219
  "dsh-live-voice.speak.segmentGap.label": "Pause between speech segments",
2085
2220
  "dsh-live-voice.speak.status.paused": "Speech paused",
@@ -2204,6 +2339,9 @@ var es = {
2204
2339
  "dsh-live-voice.recognition.commands.send": "Enviar al agente en ejecuci\xF3n",
2205
2340
  "dsh-live-voice.recognition.commands.stopSpeech": "Detener la lectura en voz alta del asistente",
2206
2341
  "dsh-live-voice.recognition.commands.title": "Comandos de voz",
2342
+ "dsh-live-voice.recognition.context.hint": "Contexto y vocabulario opcionales enviados como prompt al servicio HTTP de voz seleccionado con cada segmento de audio. Se guarda al salir del campo. La compatibilidad depende del servidor; no restringe idiomas ni garantiza el seguimiento de instrucciones. D\xE9jalo vac\xEDo para no enviar contexto.",
2343
+ "dsh-live-voice.recognition.context.label": "Contexto de transcripci\xF3n",
2344
+ "dsh-live-voice.recognition.context.placeholder": "Native Brazilian Portuguese speaker who frequently uses technical terms in English.",
2207
2345
  "dsh-live-voice.recognition.dictation.cancel": "Cancelar dictado",
2208
2346
  "dsh-live-voice.recognition.engine.label": "Motor de reconocimiento",
2209
2347
  "dsh-live-voice.recognition.headphoneMode.help": "El micr\xF3fono sigue abierto durante la reproducci\xF3n. El habla reconocida sin eco puede pausarla tras el retraso configurado; se reanuda cuando el micr\xF3fono y el audio compartido guardan silencio durante ese intervalo. La pausa manual requiere Reanudar.",
@@ -2476,7 +2614,7 @@ var es = {
2476
2614
  "dsh-live-voice.speak.rate.label": "Velocidad de habla",
2477
2615
  "dsh-live-voice.speak.responseDelay.help": "Cuando dejas de hablar, la reproducci\xF3n autom\xE1tica del asistente espera este intervalo de silencio continuo. Si vuelves a hablar, la espera se reinicia.",
2478
2616
  "dsh-live-voice.speak.responseDelay.label": "Demora de respuesta del asistente",
2479
- "dsh-live-voice.speak.responseDelay.none": "Sin retraso",
2617
+ "dsh-live-voice.speak.responseDelay.oneAndHalf": "1,5 segundos",
2480
2618
  "dsh-live-voice.speak.segmentGap.help": "Espera esta cantidad de milisegundos entre segmentos hablados consecutivos. El valor predeterminado es 400 ms.",
2481
2619
  "dsh-live-voice.speak.segmentGap.label": "Pausa entre segmentos de voz",
2482
2620
  "dsh-live-voice.speak.status.paused": "Lectura en pausa",
@@ -2601,6 +2739,9 @@ var fr = {
2601
2739
  "dsh-live-voice.recognition.commands.send": "Envoyer \xE0 l\u2019agent en cours d\u2019ex\xE9cution",
2602
2740
  "dsh-live-voice.recognition.commands.stopSpeech": "Arr\xEAter la lecture vocale de l\u2019assistant",
2603
2741
  "dsh-live-voice.recognition.commands.title": "Commandes vocales",
2742
+ "dsh-live-voice.recognition.context.hint": "Contexte et vocabulaire facultatifs envoy\xE9s comme prompt au service vocal HTTP s\xE9lectionn\xE9 avec chaque segment audio. Enregistr\xE9 \xE0 la sortie du champ. La prise en charge d\xE9pend du serveur ; ne limite pas les langues et ne garantit pas le suivi des consignes. Laissez vide pour ne pas envoyer de contexte.",
2743
+ "dsh-live-voice.recognition.context.label": "Contexte de transcription",
2744
+ "dsh-live-voice.recognition.context.placeholder": "Native Brazilian Portuguese speaker who frequently uses technical terms in English.",
2604
2745
  "dsh-live-voice.recognition.dictation.cancel": "Annuler la dict\xE9e",
2605
2746
  "dsh-live-voice.recognition.engine.label": "Moteur de reconnaissance",
2606
2747
  "dsh-live-voice.recognition.headphoneMode.help": "Le microphone reste ouvert pendant la lecture. La parole reconnue sans \xE9cho peut la mettre en pause apr\xE8s le d\xE9lai configur\xE9 ; elle reprend lorsque le microphone et le son partag\xE9 restent silencieux pendant ce d\xE9lai. Une pause manuelle n\xE9cessite Reprendre.",
@@ -2873,7 +3014,7 @@ var fr = {
2873
3014
  "dsh-live-voice.speak.rate.label": "D\xE9bit de parole",
2874
3015
  "dsh-live-voice.speak.responseDelay.help": "Apr\xE8s que vous avez cess\xE9 de parler, la lecture automatique de l\u2019assistant attend cette dur\xE9e de silence continu. Si vous reparlez, l\u2019attente recommence.",
2875
3016
  "dsh-live-voice.speak.responseDelay.label": "D\xE9lai de r\xE9ponse de l\u2019assistant",
2876
- "dsh-live-voice.speak.responseDelay.none": "Sans d\xE9lai",
3017
+ "dsh-live-voice.speak.responseDelay.oneAndHalf": "1,5 seconde",
2877
3018
  "dsh-live-voice.speak.segmentGap.help": "Attend ce nombre de millisecondes entre des segments vocaux cons\xE9cutifs. La valeur par d\xE9faut est 400 ms.",
2878
3019
  "dsh-live-voice.speak.segmentGap.label": "Pause entre les segments vocaux",
2879
3020
  "dsh-live-voice.speak.status.paused": "Lecture en pause",
@@ -2998,6 +3139,9 @@ var hi = {
2998
3139
  "dsh-live-voice.recognition.commands.send": "\u091A\u0932 \u0930\u0939\u0947 \u090F\u091C\u0947\u0902\u091F \u0915\u094B \u092D\u0947\u091C\u0947\u0902",
2999
3140
  "dsh-live-voice.recognition.commands.stopSpeech": "\u0938\u0939\u093E\u092F\u0915 \u0915\u093E \u0935\u093E\u091A\u0928 \u092C\u0902\u0926 \u0915\u0930\u0947\u0902",
3000
3141
  "dsh-live-voice.recognition.commands.title": "\u0935\u0949\u0907\u0938 \u0915\u092E\u093E\u0902\u0921",
3142
+ "dsh-live-voice.recognition.context.hint": "\u0935\u0948\u0915\u0932\u094D\u092A\u093F\u0915 \u0938\u0902\u0926\u0930\u094D\u092D \u0914\u0930 \u0936\u092C\u094D\u0926\u093E\u0935\u0932\u0940 \u0939\u0930 \u0911\u0921\u093F\u092F\u094B \u0916\u0902\u0921 \u0915\u0947 \u0938\u093E\u0925 \u091A\u0941\u0928\u0940 \u0917\u0908 HTTP \u0935\u093E\u0915\u094D \u0938\u0947\u0935\u093E \u0915\u094B prompt \u0915\u0947 \u0930\u0942\u092A \u092E\u0947\u0902 \u092D\u0947\u091C\u0947 \u091C\u093E\u0924\u0947 \u0939\u0948\u0902\u0964 \u092B\u093C\u0940\u0932\u094D\u0921 \u0938\u0947 \u092C\u093E\u0939\u0930 \u0906\u0928\u0947 \u092A\u0930 \u0938\u0939\u0947\u091C\u093E \u091C\u093E\u0924\u093E \u0939\u0948\u0964 \u0938\u092E\u0930\u094D\u0925\u0928 \u0938\u0930\u094D\u0935\u0930 \u092A\u0930 \u0928\u093F\u0930\u094D\u092D\u0930 \u0939\u0948; \u092F\u0939 \u092D\u093E\u0937\u093E\u0913\u0902 \u0915\u094B \u0938\u0940\u092E\u093F\u0924 \u0928\u0939\u0940\u0902 \u0915\u0930\u0924\u093E \u0914\u0930 \u0928\u093F\u0930\u094D\u0926\u0947\u0936\u094B\u0902 \u0915\u0947 \u092A\u093E\u0932\u0928 \u0915\u0940 \u0917\u093E\u0930\u0902\u091F\u0940 \u0928\u0939\u0940\u0902 \u0926\u0947\u0924\u093E\u0964 \u0938\u0902\u0926\u0930\u094D\u092D \u0928 \u092D\u0947\u091C\u0928\u0947 \u0915\u0947 \u0932\u093F\u090F \u0916\u093E\u0932\u0940 \u091B\u094B\u0921\u093C\u0947\u0902\u0964",
3143
+ "dsh-live-voice.recognition.context.label": "\u091F\u094D\u0930\u093E\u0902\u0938\u0915\u094D\u0930\u093F\u092A\u094D\u0936\u0928 \u0938\u0902\u0926\u0930\u094D\u092D",
3144
+ "dsh-live-voice.recognition.context.placeholder": "Native Brazilian Portuguese speaker who frequently uses technical terms in English.",
3001
3145
  "dsh-live-voice.recognition.dictation.cancel": "\u0921\u093F\u0915\u094D\u091F\u0947\u0936\u0928 \u0930\u0926\u094D\u0926 \u0915\u0930\u0947\u0902",
3002
3146
  "dsh-live-voice.recognition.engine.label": "\u0935\u093E\u0915\u094D \u092A\u0939\u091A\u093E\u0928 \u0907\u0902\u091C\u0928",
3003
3147
  "dsh-live-voice.recognition.headphoneMode.help": "\u092A\u094D\u0932\u0947\u092C\u0948\u0915 \u0915\u0947 \u0926\u094C\u0930\u093E\u0928 \u092E\u093E\u0907\u0915\u094D\u0930\u094B\u092B\u093C\u094B\u0928 \u091A\u093E\u0932\u0942 \u0930\u0939\u0924\u093E \u0939\u0948\u0964 \u092C\u093F\u0928\u093E \u092A\u094D\u0930\u0924\u093F\u0927\u094D\u0935\u0928\u093F \u0915\u0940 \u092A\u0939\u091A\u093E\u0928\u0940 \u0917\u0908 \u0906\u0935\u093E\u091C\u093C \u0928\u093F\u0930\u094D\u0927\u093E\u0930\u093F\u0924 \u0935\u093F\u0932\u0902\u092C \u0915\u0947 \u092C\u093E\u0926 \u092A\u094D\u0932\u0947\u092C\u0948\u0915 \u0930\u094B\u0915 \u0938\u0915\u0924\u0940 \u0939\u0948; \u092E\u093E\u0907\u0915\u094D\u0930\u094B\u092B\u093C\u094B\u0928 \u0914\u0930 \u0938\u093E\u091D\u093E \u092E\u0940\u091F\u093F\u0902\u0917 \u0911\u0921\u093F\u092F\u094B \u0909\u0924\u0928\u0940 \u0926\u0947\u0930 \u0936\u093E\u0902\u0924 \u0930\u0939\u0928\u0947 \u092A\u0930 \u092F\u0939 \u092B\u093F\u0930 \u091A\u0932\u0924\u093E \u0939\u0948\u0964 \u092E\u0948\u0928\u094D\u092F\u0941\u0905\u0932 \u0935\u093F\u0930\u093E\u092E \u0915\u0947 \u092C\u093E\u0926 \u092B\u093F\u0930 \u091A\u0932\u093E\u090F\u0901 \u0926\u092C\u093E\u090F\u0901\u0964",
@@ -3270,7 +3414,7 @@ var hi = {
3270
3414
  "dsh-live-voice.speak.rate.label": "\u092C\u094B\u0932\u0928\u0947 \u0915\u0940 \u0917\u0924\u093F",
3271
3415
  "dsh-live-voice.speak.responseDelay.help": "\u0906\u092A\u0915\u0947 \u092C\u094B\u0932\u0928\u093E \u092C\u0902\u0926 \u0915\u0930\u0928\u0947 \u0915\u0947 \u092C\u093E\u0926 \u0938\u0939\u093E\u092F\u0915 \u0915\u093E \u0938\u094D\u0935\u091A\u093E\u0932\u093F\u0924 \u092A\u094D\u0932\u0947\u092C\u0948\u0915 \u0907\u0924\u0928\u0940 \u0926\u0947\u0930 \u0924\u0915 \u0932\u0917\u093E\u0924\u093E\u0930 \u092E\u094C\u0928 \u0930\u0939\u0928\u0947 \u0915\u0940 \u092A\u094D\u0930\u0924\u0940\u0915\u094D\u0937\u093E \u0915\u0930\u0924\u093E \u0939\u0948\u0964 \u0926\u094B\u092C\u093E\u0930\u093E \u092C\u094B\u0932\u0928\u0947 \u092A\u0930 \u092A\u094D\u0930\u0924\u0940\u0915\u094D\u0937\u093E \u092B\u093F\u0930 \u0938\u0947 \u0936\u0941\u0930\u0942 \u0939\u094B\u0924\u0940 \u0939\u0948\u0964",
3272
3416
  "dsh-live-voice.speak.responseDelay.label": "\u0938\u0939\u093E\u092F\u0915 \u0915\u0947 \u091C\u0935\u093E\u092C \u0915\u093E \u0935\u093F\u0932\u0902\u092C",
3273
- "dsh-live-voice.speak.responseDelay.none": "\u0915\u094B\u0908 \u0935\u093F\u0932\u0902\u092C \u0928\u0939\u0940\u0902",
3417
+ "dsh-live-voice.speak.responseDelay.oneAndHalf": "1.5 \u0938\u0947\u0915\u0902\u0921",
3274
3418
  "dsh-live-voice.speak.segmentGap.help": "\u0932\u0917\u093E\u0924\u093E\u0930 \u092C\u094B\u0932\u0947 \u0917\u090F \u0939\u093F\u0938\u094D\u0938\u094B\u0902 \u0915\u0947 \u092C\u0940\u091A \u0907\u0924\u0928\u0947 \u092E\u093F\u0932\u0940\u0938\u0947\u0915\u0902\u0921 \u0930\u0941\u0915\u0947\u0902\u0964 \u0921\u093F\u092B\u093C\u0949\u0932\u094D\u091F 400 ms \u0939\u0948\u0964",
3275
3419
  "dsh-live-voice.speak.segmentGap.label": "\u092C\u094B\u0932\u0947 \u0917\u090F \u0939\u093F\u0938\u094D\u0938\u094B\u0902 \u0915\u0947 \u092C\u0940\u091A \u0935\u093F\u0930\u093E\u092E",
3276
3420
  "dsh-live-voice.speak.status.paused": "\u0935\u093E\u091A\u0928 \u0920\u0939\u0930\u093E \u0939\u0941\u0906 \u0939\u0948",
@@ -3395,6 +3539,9 @@ var ptBR = {
3395
3539
  "dsh-live-voice.recognition.commands.send": "Enviar ao agente em execu\xE7\xE3o",
3396
3540
  "dsh-live-voice.recognition.commands.stopSpeech": "Parar fala do assistente",
3397
3541
  "dsh-live-voice.recognition.commands.title": "Comandos de voz",
3542
+ "dsh-live-voice.recognition.context.hint": "Contexto e vocabul\xE1rio opcionais enviados como prompt ao servi\xE7o HTTP de voz selecionado, junto com cada trecho de \xE1udio. Salvo ao sair do campo. O suporte depende do servidor; n\xE3o restringe idiomas nem garante que instru\xE7\xF5es sejam seguidas. Deixe vazio para n\xE3o enviar contexto.",
3543
+ "dsh-live-voice.recognition.context.label": "Contexto da transcri\xE7\xE3o",
3544
+ "dsh-live-voice.recognition.context.placeholder": "Native Brazilian Portuguese speaker who frequently uses technical terms in English.",
3398
3545
  "dsh-live-voice.recognition.dictation.cancel": "Cancelar ditado",
3399
3546
  "dsh-live-voice.recognition.engine.label": "Mecanismo de reconhecimento",
3400
3547
  "dsh-live-voice.recognition.headphoneMode.help": "O microfone fica aberto durante a reprodu\xE7\xE3o. Fala reconhecida sem eco pode pausar ap\xF3s o atraso configurado; a reprodu\xE7\xE3o retoma quando microfone e \xE1udio compartilhado da reuni\xE3o ficam em sil\xEAncio por esse intervalo. Pausa manual exige Retomar.",
@@ -3667,7 +3814,7 @@ var ptBR = {
3667
3814
  "dsh-live-voice.speak.rate.label": "Velocidade da fala",
3668
3815
  "dsh-live-voice.speak.responseDelay.help": "Depois que voc\xEA parar de falar, a reprodu\xE7\xE3o autom\xE1tica do assistente aguardar\xE1 esse tempo de sil\xEAncio cont\xEDnuo. Falar novamente reinicia a contagem.",
3669
3816
  "dsh-live-voice.speak.responseDelay.label": "Atraso da resposta do assistente",
3670
- "dsh-live-voice.speak.responseDelay.none": "Sem atraso",
3817
+ "dsh-live-voice.speak.responseDelay.oneAndHalf": "1,5 segundo",
3671
3818
  "dsh-live-voice.speak.segmentGap.help": "Aguarde esta quantidade de milissegundos entre trechos falados consecutivos. O padr\xE3o \xE9 400 ms.",
3672
3819
  "dsh-live-voice.speak.segmentGap.label": "Pausa entre trechos de fala",
3673
3820
  "dsh-live-voice.speak.status.paused": "Fala pausada",
@@ -3792,6 +3939,9 @@ var zh = {
3792
3939
  "dsh-live-voice.recognition.commands.send": "\u53D1\u9001\u7ED9\u8FD0\u884C\u4E2D\u7684\u667A\u80FD\u4F53",
3793
3940
  "dsh-live-voice.recognition.commands.stopSpeech": "\u505C\u6B62\u52A9\u624B\u6717\u8BFB",
3794
3941
  "dsh-live-voice.recognition.commands.title": "\u8BED\u97F3\u547D\u4EE4",
3942
+ "dsh-live-voice.recognition.context.hint": "\u53EF\u9009\u7684\u80CC\u666F\u548C\u8BCD\u6C47\u4F1A\u968F\u6BCF\u6BB5\u97F3\u9891\u4EE5 prompt \u5B57\u6BB5\u53D1\u9001\u5230\u6240\u9009 HTTP \u8BED\u97F3\u670D\u52A1\u3002\u79BB\u5F00\u8F93\u5165\u6846\u65F6\u4FDD\u5B58\u3002\u652F\u6301\u60C5\u51B5\u53D6\u51B3\u4E8E\u670D\u52A1\u5668\uFF1B\u4E0D\u4F1A\u9650\u5236\u8BED\u8A00\uFF0C\u4E5F\u4E0D\u4FDD\u8BC1\u9075\u5FAA\u6307\u4EE4\u3002\u7559\u7A7A\u5219\u4E0D\u53D1\u9001\u4E0A\u4E0B\u6587\u3002",
3943
+ "dsh-live-voice.recognition.context.label": "\u8F6C\u5199\u4E0A\u4E0B\u6587",
3944
+ "dsh-live-voice.recognition.context.placeholder": "Native Brazilian Portuguese speaker who frequently uses technical terms in English.",
3795
3945
  "dsh-live-voice.recognition.dictation.cancel": "\u53D6\u6D88\u542C\u5199",
3796
3946
  "dsh-live-voice.recognition.engine.label": "\u8BC6\u522B\u5F15\u64CE",
3797
3947
  "dsh-live-voice.recognition.headphoneMode.help": "\u64AD\u653E\u65F6\u9EA6\u514B\u98CE\u4FDD\u6301\u5F00\u542F\u3002\u8BC6\u522B\u5230\u975E\u56DE\u58F0\u8BED\u97F3\u540E\uFF0C\u53EF\u5728\u8BBE\u5B9A\u5EF6\u8FDF\u540E\u6682\u505C\u64AD\u653E\uFF1B\u9EA6\u514B\u98CE\u548C\u5171\u4EAB\u4F1A\u8BAE\u97F3\u9891\u6301\u7EED\u9759\u97F3\u8FBE\u5230\u8BE5\u65F6\u957F\u540E\u6062\u590D\u3002\u624B\u52A8\u6682\u505C\u540E\u9700\u70B9\u51FB\u6062\u590D\u3002",
@@ -4064,7 +4214,7 @@ var zh = {
4064
4214
  "dsh-live-voice.speak.rate.label": "\u8BED\u901F",
4065
4215
  "dsh-live-voice.speak.responseDelay.help": "\u4F60\u505C\u6B62\u8BF4\u8BDD\u540E\uFF0C\u52A9\u624B\u4F1A\u7B49\u5F85\u6B64\u65F6\u957F\u7684\u8FDE\u7EED\u9759\u97F3\uFF0C\u7136\u540E\u81EA\u52A8\u64AD\u653E\u8BED\u97F3\u3002\u518D\u6B21\u8BF4\u8BDD\u4F1A\u91CD\u65B0\u5F00\u59CB\u7B49\u5F85\u3002",
4066
4216
  "dsh-live-voice.speak.responseDelay.label": "\u52A9\u624B\u54CD\u5E94\u5EF6\u8FDF",
4067
- "dsh-live-voice.speak.responseDelay.none": "\u65E0\u5EF6\u8FDF",
4217
+ "dsh-live-voice.speak.responseDelay.oneAndHalf": "1.5 \u79D2",
4068
4218
  "dsh-live-voice.speak.segmentGap.help": "\u5728\u8FDE\u7EED\u8BED\u97F3\u7247\u6BB5\u4E4B\u95F4\u7B49\u5F85\u6B64\u6BEB\u79D2\u6570\u3002\u9ED8\u8BA4\u503C\u4E3A 400 ms\u3002",
4069
4219
  "dsh-live-voice.speak.segmentGap.label": "\u8BED\u97F3\u7247\u6BB5\u4E4B\u95F4\u7684\u505C\u987F",
4070
4220
  "dsh-live-voice.speak.status.paused": "\u8BED\u97F3\u5DF2\u6682\u505C",
@@ -5062,6 +5212,19 @@ function SpeechStatusBar({ controller }) {
5062
5212
  ));
5063
5213
  }
5064
5214
 
5215
+ // src/modules/recognition/transcriptionContext.ts
5216
+ var TRANSCRIPTION_CONTEXT_MAX_LENGTH = 1e3;
5217
+ function transcriptionContext(value) {
5218
+ if (value === void 0 || value === null) return "";
5219
+ if (typeof value !== "string" || value.length > TRANSCRIPTION_CONTEXT_MAX_LENGTH || value.includes("\0"))
5220
+ throw new Error("Transcription context must contain at most 1000 characters without NUL.");
5221
+ return value.trim();
5222
+ }
5223
+ function transcriptionContextHeader(value) {
5224
+ const text = transcriptionContext(value);
5225
+ return text ? { "x-dlv-transcription-context": encodeURIComponent(text) } : {};
5226
+ }
5227
+
5065
5228
  // src/modules/core/speechDefaults.ts
5066
5229
  var notices = {
5067
5230
  en: "You can see the code in our conversation history.",
@@ -5119,6 +5282,7 @@ var defaultQwenVoice = qwenVoices[0].value;
5119
5282
  var defaultAgentVoiceContext = `Live Voice output is active. Your entire user-facing response will be spoken aloud.
5120
5283
  Be concise and conversational. Lead with the answer or next action. Avoid unnecessary repetition, long preambles, dense lists, raw code, paths, and verbose status narration.
5121
5284
  Do not narrate routine tool activity by default. If the user explicitly asks you to keep them informed while working, provide brief spoken progress updates only at meaningful milestones.`;
5285
+ var assistantSpeechDelayOptions = Object.freeze([0.6, 1.5, 2, 3, 4, 5]);
5122
5286
  var autoSendDelayOptions = Object.freeze([0.6, 0.8, 1, 2, 3, 4, 5, 6]);
5123
5287
  var recordingBitrateOptions = Object.freeze([24e3, 32e3, 48e3, 64e3, 96e3]);
5124
5288
  var defaultSettings = Object.freeze({
@@ -5154,10 +5318,11 @@ var defaultSettings = Object.freeze({
5154
5318
  sendingMode: "manual",
5155
5319
  dshBusyEnterBehavior: "queue",
5156
5320
  autoSendDelaySeconds: 4,
5157
- assistantSpeechDelaySeconds: 0,
5321
+ assistantSpeechDelaySeconds: 1.5,
5158
5322
  mode: "speaker",
5159
5323
  lang: "pt-BR",
5160
5324
  recognitionLang: "pt-BR",
5325
+ recognitionContext: "",
5161
5326
  voice: "",
5162
5327
  inputDeviceId: "",
5163
5328
  outputDeviceId: "",
@@ -5219,6 +5384,7 @@ function normalizeSettings(value) {
5219
5384
  browserQwenModel: "qwen-0.6b",
5220
5385
  browserAsrDevice: source.browserAsrDevice === "wasm" ? "wasm" : "webgpu",
5221
5386
  recognitionProcessLocally: typeof source.recognitionProcessLocally === "boolean" ? source.recognitionProcessLocally : defaultSettings.recognitionProcessLocally,
5387
+ recognitionContext: typeof source.recognitionContext === "string" ? source.recognitionContext.replaceAll("\0", "").slice(0, TRANSCRIPTION_CONTEXT_MAX_LENGTH).trim() : defaultSettings.recognitionContext,
5222
5388
  recognitionAutoInstall: typeof source.recognitionAutoInstall === "boolean" ? source.recognitionAutoInstall : defaultSettings.recognitionAutoInstall,
5223
5389
  voiceDetectionEngine: ["energy", "silero"].includes(source.voiceDetectionEngine) ? source.voiceDetectionEngine : defaultSettings.voiceDetectionEngine,
5224
5390
  voiceDetectionPreset: source.voiceDetectionPreset === "custom" || Object.hasOwn(voiceDetectionPresets, source.voiceDetectionPreset) ? source.voiceDetectionPreset : defaultSettings.voiceDetectionPreset,
@@ -5256,7 +5422,9 @@ function normalizeSettings(value) {
5256
5422
  dshBusyEnterBehavior: ["queue", "steer"].includes(source.dshBusyEnterBehavior) ? source.dshBusyEnterBehavior : defaultSettings.dshBusyEnterBehavior,
5257
5423
  sendingMode: source.sendingMode === "automatic" ? "queue" : ["manual", "queue", "steer"].includes(source.sendingMode) ? source.sendingMode : defaultSettings.sendingMode,
5258
5424
  autoSendDelaySeconds: autoSendDelayOptions.includes(source.autoSendDelaySeconds) ? source.autoSendDelaySeconds : defaultSettings.autoSendDelaySeconds,
5259
- assistantSpeechDelaySeconds: Number.isInteger(source.assistantSpeechDelaySeconds) && source.assistantSpeechDelaySeconds >= 0 && source.assistantSpeechDelaySeconds <= 4 ? source.assistantSpeechDelaySeconds : defaultSettings.assistantSpeechDelaySeconds,
5425
+ assistantSpeechDelaySeconds: assistantSpeechDelayOptions.includes(
5426
+ source.assistantSpeechDelaySeconds
5427
+ ) ? source.assistantSpeechDelaySeconds : defaultSettings.assistantSpeechDelaySeconds,
5260
5428
  mode: ["speaker", "headphones"].includes(source.mode) ? source.mode : defaultSettings.mode,
5261
5429
  lang: typeof source.lang === "string" && /^[a-z]{2,3}(?:-[A-Za-z0-9]{2,8})*$/.test(source.lang) ? source.lang : defaultSettings.lang,
5262
5430
  recognitionLang: source.recognitionLang === "auto" || typeof source.recognitionLang === "string" && /^[a-z]{2,3}(?:-[A-Za-z0-9]{2,8})*$/.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,
@@ -6881,8 +7049,23 @@ function cancellable(promise, signal) {
6881
7049
  });
6882
7050
  }
6883
7051
  var VoiceCoordinator = class _VoiceCoordinator {
6884
- constructor({ recognition, engines, meter, composer, settings = {}, persistSettings }) {
6885
- Object.assign(this, { recognition, engines, meter, composer, persistSettings });
7052
+ constructor({
7053
+ recognition,
7054
+ engines,
7055
+ meter,
7056
+ composer,
7057
+ settings = {},
7058
+ persistSettings,
7059
+ recordingTimeline
7060
+ }) {
7061
+ Object.assign(this, {
7062
+ recognition,
7063
+ engines,
7064
+ meter,
7065
+ composer,
7066
+ persistSettings,
7067
+ recordingTimeline
7068
+ });
6886
7069
  this.listeners = /* @__PURE__ */ new Set();
6887
7070
  this.transcript = new TranscriptDraft();
6888
7071
  this.epoch = 0;
@@ -7696,6 +7879,7 @@ var VoiceCoordinator = class _VoiceCoordinator {
7696
7879
  this.queue = [];
7697
7880
  }
7698
7881
  async stopSpeech(resumeListening = true) {
7882
+ this.activeRecordingTimeline = null;
7699
7883
  this._cancelAssistantSpeechTimer();
7700
7884
  this.assistantSpeechNotBefore = 0;
7701
7885
  this._cancelInterruptionResume();
@@ -7803,8 +7987,10 @@ var VoiceCoordinator = class _VoiceCoordinator {
7803
7987
  try {
7804
7988
  const result = await this.engines[this.snapshot.speechEngine || this.snapshot.settings.engine].pause();
7805
7989
  if (result !== false) this._cancelSpeechPrefetch();
7806
- if (epoch === this.speechEpoch && control === this.controlEpoch && this.snapshot.speaking && result !== false)
7990
+ if (epoch === this.speechEpoch && control === this.controlEpoch && this.snapshot.speaking && result !== false) {
7807
7991
  this.patch({ paused: true });
7992
+ this.activeRecordingTimeline?.("assistant-pause");
7993
+ }
7808
7994
  } catch (error) {
7809
7995
  if (epoch === this.speechEpoch) this.patch({ error: message(error) });
7810
7996
  }
@@ -7844,6 +8030,7 @@ var VoiceCoordinator = class _VoiceCoordinator {
7844
8030
  const result = await this.engines[this.snapshot.speechEngine || this.snapshot.settings.engine].resume();
7845
8031
  if (epoch === this.speechEpoch && control === this.controlEpoch && result !== false) {
7846
8032
  this.patch({ paused: false });
8033
+ this.activeRecordingTimeline?.("assistant-resume");
7847
8034
  this._prefetchSpeech();
7848
8035
  }
7849
8036
  } catch (error) {
@@ -7887,6 +8074,10 @@ var VoiceCoordinator = class _VoiceCoordinator {
7887
8074
  const controller = new AbortController();
7888
8075
  this.speechController = controller;
7889
8076
  this.speechLaunchPending = true;
8077
+ const timeline = this.activeRecordingTimeline ?? this.recordingTimeline?.();
8078
+ this.activeRecordingTimeline = timeline;
8079
+ let playbackStarted = false;
8080
+ let playbackFinished = false;
7890
8081
  const engine = this.engines[this.snapshot.settings.engine];
7891
8082
  this._resetStaleSpeechItem(item, this._speechPreparationKey());
7892
8083
  item.engine = engine;
@@ -7911,6 +8102,7 @@ var VoiceCoordinator = class _VoiceCoordinator {
7911
8102
  if (this.speechStopError) throw this.speechStopError;
7912
8103
  if (epoch !== this.speechEpoch || this.disposed) return;
7913
8104
  if (this.snapshot.settings.mode === "speaker") {
8105
+ if (this.snapshot.listening) timeline?.("capture-gated");
7914
8106
  await this.stopListening();
7915
8107
  if (this.inputReleaseError) throw this.inputReleaseError;
7916
8108
  }
@@ -7921,7 +8113,11 @@ var VoiceCoordinator = class _VoiceCoordinator {
7921
8113
  if (queued?.prepareError) throw queued.prepareError;
7922
8114
  const options = {
7923
8115
  onPlaybackStart: () => {
7924
- if (epoch === this.speechEpoch && !this.disposed) this.patch({ speechLoading: false });
8116
+ if (epoch === this.speechEpoch && !this.disposed && !controller.signal.aborted && !playbackFinished) {
8117
+ if (!playbackStarted) timeline?.("assistant-start");
8118
+ playbackStarted = true;
8119
+ this.patch({ speechLoading: false });
8120
+ }
7925
8121
  },
7926
8122
  voice: this.snapshot.settings.voice || void 0,
7927
8123
  rate: this.snapshot.settings.rate,
@@ -7954,6 +8150,10 @@ var VoiceCoordinator = class _VoiceCoordinator {
7954
8150
  if (error.name !== "AbortError") this.patch({ error: message(error) });
7955
8151
  }
7956
8152
  } finally {
8153
+ playbackFinished = true;
8154
+ if (playbackStarted) timeline?.("assistant-end");
8155
+ if (this.activeRecordingTimeline === timeline && epoch === this.speechEpoch && !this.queue.length)
8156
+ this.activeRecordingTimeline = null;
7957
8157
  if (epoch === this.speechEpoch && !this.disposed) {
7958
8158
  this.patch({
7959
8159
  speaking: false,
@@ -9527,6 +9727,12 @@ var PinDetector = class {
9527
9727
  }
9528
9728
  } else this.candidate = 0;
9529
9729
  }
9730
+ /** Advance past unclassified saved samples without inventing a speech/silence transition. */
9731
+ skipTo(position) {
9732
+ this.at = Math.max(this.at, position);
9733
+ this.candidate = 0;
9734
+ this.quiet = 0;
9735
+ }
9530
9736
  finish() {
9531
9737
  if (this.speaking) {
9532
9738
  this.speaking = false;
@@ -9538,6 +9744,7 @@ var PinDetector = class {
9538
9744
 
9539
9745
  // src/modules/recording/RecordingRecognition.ts
9540
9746
  var asError = (value) => value instanceof Error ? value : new Error(String(value));
9747
+ var TRANSCRIPTION_BOUNDARY_MARGIN_MS = 800;
9541
9748
  var uid = () => crypto.randomUUID();
9542
9749
  async function startRecordingRecognition(engine, options, capture) {
9543
9750
  const { signal, lang = engine.lang, onResult, onActivity, onError, onProcessingChange } = options;
@@ -9546,8 +9753,8 @@ async function startRecordingRecognition(engine, options, capture) {
9546
9753
  enabled: engine.recordingEnabled !== false,
9547
9754
  keepPrevious: engine.recordingKeepPrevious ?? 1,
9548
9755
  bitrate: engine.recordingBitrate ?? 48e3,
9549
- preRollMs: engine.recordingPreRollMs ?? 320,
9550
- postRollMs: engine.recordingPostRollMs ?? 320
9756
+ preRollMs: Math.max(TRANSCRIPTION_BOUNDARY_MARGIN_MS, engine.recordingPreRollMs ?? 320),
9757
+ postRollMs: Math.max(TRANSCRIPTION_BOUNDARY_MARGIN_MS, engine.recordingPostRollMs ?? 320)
9551
9758
  };
9552
9759
  const generation = engine.startGeneration;
9553
9760
  const writer = await engine.recordings.begin({
@@ -9608,20 +9815,22 @@ async function startRecordingRecognition(engine, options, capture) {
9608
9815
  };
9609
9816
  engine.session = session;
9610
9817
  const valid = () => engine.session === session && !signal?.aborted;
9818
+ const awaitingTail = [];
9611
9819
  let stopping = false;
9820
+ let recoveryPending = false;
9612
9821
  let captureFailure;
9613
9822
  const captureValid = () => engine.session === session || stopping;
9614
9823
  const notify2 = () => {
9615
9824
  if (valid())
9616
9825
  session.onProcessingChange?.({
9617
- queued: session.transcriptionQueue.length,
9826
+ queued: awaitingTail.length + session.transcriptionQueue.length,
9618
9827
  active: !!session.activeRequest,
9619
9828
  // Only queued/running speech attempts belong in the visible transcription status.
9620
- transcribing: session.transcriptionQueue.length + (session.activeRequest ? 1 : 0),
9829
+ transcribing: awaitingTail.length + session.transcriptionQueue.length + (session.activeRequest ? 1 : 0),
9621
9830
  // Speech-boundary bookkeeping must settle before delivery, but silent PCM writes must not reset it.
9622
- deliveryPending: session.transcriptionQueue.length + (session.activeRequest ? 1 : 0) + session.pendingSpeechWrites + (session.finishing ? 1 : 0),
9831
+ deliveryPending: awaitingTail.length + session.transcriptionQueue.length + (session.activeRequest ? 1 : 0) + session.pendingSpeechWrites + (session.finishing || recoveryPending ? 1 : 0),
9623
9832
  // All writes still protect release, playback admission and page leave.
9624
- pending: session.transcriptionQueue.length + (session.activeRequest ? 1 : 0) + session.pendingWrites + (session.finishing ? 1 : 0)
9833
+ pending: awaitingTail.length + session.transcriptionQueue.length + (session.activeRequest ? 1 : 0) + session.pendingWrites + (session.finishing || recoveryPending ? 1 : 0)
9625
9834
  });
9626
9835
  };
9627
9836
  const failure = (error) => {
@@ -9672,10 +9881,7 @@ async function startRecordingRecognition(engine, options, capture) {
9672
9881
  const queueSpeech = async (pin2, end) => {
9673
9882
  for (let at2 = Math.max(pin2.atSample, queuedThrough); at2 < end; at2 += max) {
9674
9883
  const requestedStart = Math.max(0, at2 - policy.preRollMs * rate / 1e3);
9675
- const requestedEnd = Math.min(
9676
- writer.session.capturedSamples,
9677
- Math.min(end, at2 + max) + policy.postRollMs * rate / 1e3
9678
- );
9884
+ const requestedEnd = Math.min(end, at2 + max) + policy.postRollMs * rate / 1e3;
9679
9885
  const attempt = {
9680
9886
  id: uid(),
9681
9887
  segmentId: pin2.id,
@@ -9687,13 +9893,26 @@ async function startRecordingRecognition(engine, options, capture) {
9687
9893
  createdAt: (/* @__PURE__ */ new Date()).toISOString(),
9688
9894
  status: "pending"
9689
9895
  };
9896
+ awaitingTail.push(attempt);
9897
+ notify2();
9898
+ }
9899
+ queuedThrough = Math.max(queuedThrough, end);
9900
+ await releaseCapturedRanges();
9901
+ };
9902
+ async function releaseCapturedRanges() {
9903
+ if (!valid()) return;
9904
+ const captured = writer.session.capturedSamples;
9905
+ while (awaitingTail.length && (session.finished || awaitingTail[0].endSample <= captured)) {
9906
+ const pending = awaitingTail[0];
9907
+ const attempt = { ...pending, endSample: Math.min(pending.endSample, captured) };
9690
9908
  await writer.attempt(attempt);
9909
+ if (!valid()) return;
9910
+ awaitingTail.shift();
9691
9911
  session.transcriptionQueue.push(attempt);
9692
9912
  notify2();
9693
9913
  }
9694
- queuedThrough = Math.max(queuedThrough, end);
9695
9914
  void drain();
9696
- };
9915
+ }
9697
9916
  const pin = (kind, at2, confirmed) => {
9698
9917
  const startSample = Math.min(
9699
9918
  session.capturedSamples,
@@ -9760,7 +9979,8 @@ async function startRecordingRecognition(engine, options, capture) {
9760
9979
  "content-type": "audio/wav",
9761
9980
  "x-dlv-client-id": session.operation,
9762
9981
  "x-dlv-operation-id": attempt.id,
9763
- "x-dlv-language": lang
9982
+ "x-dlv-language": lang,
9983
+ ...transcriptionContextHeader(engine.recognitionContext)
9764
9984
  },
9765
9985
  credentials: "same-origin",
9766
9986
  body: capture.encode(pcm, rate),
@@ -9814,7 +10034,114 @@ async function startRecordingRecognition(engine, options, capture) {
9814
10034
  return;
9815
10035
  }
9816
10036
  const copy = pcm.slice();
9817
- void serialize(() => writer.append(copy), true).catch(() => {
10037
+ void serialize(async () => {
10038
+ await writer.append(copy);
10039
+ await releaseCapturedRanges();
10040
+ }, true).catch(() => {
10041
+ });
10042
+ };
10043
+ let vadGeneration = 0;
10044
+ let restartAttempts = 0;
10045
+ let recovering = null;
10046
+ let deferredFailure = null;
10047
+ const sileroValid = () => valid() && !stopping && !session.finished && !session.finishing;
10048
+ const releaseSilero = () => {
10049
+ ++vadGeneration;
10050
+ const previousCapture = session.capture;
10051
+ const previousVad = session.vad;
10052
+ session.capture = null;
10053
+ session.vad = null;
10054
+ try {
10055
+ previousCapture?.release();
10056
+ } finally {
10057
+ previousVad?.release();
10058
+ }
10059
+ };
10060
+ const terminalVadFailure = (error) => {
10061
+ if (!valid() || stopping || session.finished) return;
10062
+ captureFailure = asError(error).message;
10063
+ onError?.(new Error(engine.vadErrorMessage(error)));
10064
+ void engine.stop();
10065
+ };
10066
+ const installSilero = async () => {
10067
+ const token = ++vadGeneration;
10068
+ let installing = true;
10069
+ let installationFailure;
10070
+ const failed = (error) => {
10071
+ if (token !== vadGeneration || !sileroValid()) return;
10072
+ if (installing) installationFailure = error;
10073
+ else recoverSilero(error);
10074
+ };
10075
+ const vad = engine.vadFactory({
10076
+ onProbability: ({ probability, pcm }) => {
10077
+ if (token !== vadGeneration || !captureValid() || session.finished) return;
10078
+ restartAttempts = 0;
10079
+ detector.frame(probability, pcm.length);
10080
+ checkLiveSplit();
10081
+ },
10082
+ onError: failed,
10083
+ onReset: () => failed(new Error("VAD analysis interrupted."))
10084
+ });
10085
+ session.vad = vad;
10086
+ try {
10087
+ const attached = await engine.vadCapture(
10088
+ engine.meter.context,
10089
+ engine.meter.source,
10090
+ vad,
10091
+ (pcm, sampleRate) => {
10092
+ if (token === vadGeneration) ingest(pcm, sampleRate);
10093
+ }
10094
+ );
10095
+ if (token !== vadGeneration || !sileroValid()) {
10096
+ attached.release();
10097
+ vad.release();
10098
+ throw Object.assign(new Error("Cancelled"), { name: "AbortError" });
10099
+ }
10100
+ session.capture = attached;
10101
+ if (installationFailure) throw installationFailure;
10102
+ } finally {
10103
+ installing = false;
10104
+ }
10105
+ };
10106
+ const retrySilero = async (initialError) => {
10107
+ let error = initialError;
10108
+ while (sileroValid()) {
10109
+ if (error) {
10110
+ releaseSilero();
10111
+ if (restartAttempts >= 2) throw error;
10112
+ restartAttempts++;
10113
+ detector.skipTo(Math.round(session.capturedSamples * 16e3 / rate));
10114
+ await Promise.resolve();
10115
+ }
10116
+ if (!sileroValid()) return;
10117
+ try {
10118
+ await installSilero();
10119
+ return;
10120
+ } catch (failure2) {
10121
+ if (!sileroValid() || asError(failure2).name === "AbortError") return;
10122
+ error = failure2;
10123
+ }
10124
+ }
10125
+ };
10126
+ const recoverSilero = (error) => {
10127
+ if (!sileroValid()) return;
10128
+ if (recovering) {
10129
+ ++vadGeneration;
10130
+ deferredFailure = error;
10131
+ return;
10132
+ }
10133
+ ++vadGeneration;
10134
+ recoveryPending = true;
10135
+ notify2();
10136
+ recovering = Promise.resolve().then(() => retrySilero(error)).catch(terminalVadFailure).finally(() => {
10137
+ recovering = null;
10138
+ const pending = deferredFailure;
10139
+ deferredFailure = null;
10140
+ if (pending) recoverSilero(pending);
10141
+ else {
10142
+ recoveryPending = false;
10143
+ notify2();
10144
+ }
9818
10145
  });
9819
10146
  };
9820
10147
  session.finishRecording = () => {
@@ -9822,10 +10149,12 @@ async function startRecordingRecognition(engine, options, capture) {
9822
10149
  session.finishing = true;
9823
10150
  session.finishPromise = Promise.resolve().then(async () => {
9824
10151
  try {
10152
+ await recovering;
9825
10153
  await session.capture?.finish();
9826
10154
  if (!valid()) return;
9827
10155
  session.finished = true;
9828
10156
  detector.finish();
10157
+ await serialize(releaseCapturedRanges);
9829
10158
  await session.tail;
9830
10159
  await writer.finish();
9831
10160
  session.vad?.release();
@@ -9843,6 +10172,8 @@ async function startRecordingRecognition(engine, options, capture) {
9843
10172
  };
9844
10173
  session.stopRecording = async () => {
9845
10174
  stopping = true;
10175
+ awaitingTail.length = 0;
10176
+ await recovering;
9846
10177
  session.activeRequest?.abort();
9847
10178
  if (session.finishing) session.capture?.release();
9848
10179
  else {
@@ -9872,31 +10203,7 @@ async function startRecordingRecognition(engine, options, capture) {
9872
10203
  signal?.addEventListener("abort", session.abort, { once: true });
9873
10204
  try {
9874
10205
  if (capture.silero) {
9875
- session.vad = engine.vadFactory({
9876
- onProbability: ({ probability, pcm }) => {
9877
- if (captureValid() && !session.finished) {
9878
- detector.frame(probability, pcm.length);
9879
- checkLiveSplit();
9880
- }
9881
- },
9882
- onError: (error) => {
9883
- captureFailure = error.message;
9884
- if (valid()) onError?.(new Error(engine.vadErrorMessage(error)));
9885
- void engine.stop();
9886
- },
9887
- onReset: () => {
9888
- captureFailure = "VAD analysis interrupted.";
9889
- if (valid())
9890
- onError?.(new Error(engine.vadErrorMessage(new Error("VAD analysis interrupted."))));
9891
- void engine.stop();
9892
- }
9893
- });
9894
- session.capture = await engine.vadCapture(
9895
- engine.meter.context,
9896
- engine.meter.source,
9897
- session.vad,
9898
- ingest
9899
- );
10206
+ await retrySilero();
9900
10207
  } else {
9901
10208
  const processor = engine.meter.context.createScriptProcessor(4096, 1, 1);
9902
10209
  const gain = engine.meter.context.createGain();
@@ -9982,6 +10289,7 @@ var WhisperHttpRecognitionEngine = class {
9982
10289
  voiceDetectionPreset = defaultSettings.voiceDetectionPreset,
9983
10290
  voiceDetectionCustomSilenceMs = defaultSettings.voiceDetectionCustomSilenceMs,
9984
10291
  voiceDetectionCustomParagraphMs = defaultSettings.voiceDetectionCustomParagraphMs,
10292
+ recognitionContext = "",
9985
10293
  maxUtteranceSeconds = 60,
9986
10294
  recordings,
9987
10295
  inputSource = "microphone",
@@ -10008,6 +10316,7 @@ var WhisperHttpRecognitionEngine = class {
10008
10316
  this.voiceDetectionPreset = voiceDetectionPreset;
10009
10317
  this.voiceDetectionCustomSilenceMs = voiceDetectionCustomSilenceMs;
10010
10318
  this.voiceDetectionCustomParagraphMs = voiceDetectionCustomParagraphMs;
10319
+ this.recognitionContext = recognitionContext;
10011
10320
  this.maxUtteranceSeconds = maxUtteranceSeconds;
10012
10321
  Object.assign(this, {
10013
10322
  recordings: recordings ?? createRecordingRepository(),
@@ -10584,8 +10893,8 @@ function createDeviceOptions(devices, kind, systemDefault, fallback) {
10584
10893
  var import_react36 = __toESM(require("react"), 1);
10585
10894
 
10586
10895
  // src/modules/settings/services/releases.ts
10587
- var CURRENT_VERSION = "0.6.1";
10588
- var TESTED_DSH_VERSION = "0.2.1-alpha.3";
10896
+ var CURRENT_VERSION = "0.6.2";
10897
+ var TESTED_DSH_VERSION = "0.2.1-rc.2";
10589
10898
  var REPOSITORY_URL = "https://github.com/victorwads/dsh-live-voice";
10590
10899
  var RELEASES_URL = `${REPOSITORY_URL}/releases`;
10591
10900
  var LATEST_RELEASE_API_URL = "https://api.github.com/repos/victorwads/dsh-live-voice/releases/latest";
@@ -10813,7 +11122,6 @@ function VoiceModeSettings({
10813
11122
 
10814
11123
  // src/modules/settings/sections/conversation/ConversationDelaySettings.tsx
10815
11124
  var import_react40 = __toESM(require("react"), 1);
10816
- var DELAYS = [0, 1, 2, 3, 4];
10817
11125
  function ConversationDelaySettings({ settings, updateSettings }) {
10818
11126
  const { scoped: commons } = useLanguage((ctx) => ctx.commons);
10819
11127
  const { scoped: speak } = useLanguage((ctx) => ctx.speak);
@@ -10822,10 +11130,12 @@ function ConversationDelaySettings({ settings, updateSettings }) {
10822
11130
  {
10823
11131
  label: speak.responseDelay.label(),
10824
11132
  hint: speak.responseDelay.help(),
10825
- value: String(settings.assistantSpeechDelaySeconds ?? 0),
10826
- options: DELAYS.map((seconds) => ({
11133
+ value: String(
11134
+ settings.assistantSpeechDelaySeconds ?? defaultSettings.assistantSpeechDelaySeconds
11135
+ ),
11136
+ options: assistantSpeechDelayOptions.map((seconds) => ({
10827
11137
  value: String(seconds),
10828
- label: seconds === 0 ? speak.responseDelay.none() : seconds === 1 ? commons.second() : commons.seconds({ seconds })
11138
+ label: seconds < 1 ? commons.milliseconds({ milliseconds: seconds * 1e3 }) : seconds === 1.5 ? speak.responseDelay.oneAndHalf() : commons.seconds({ seconds })
10829
11139
  })),
10830
11140
  onChange: (event) => updateSettings({ assistantSpeechDelaySeconds: Number(event.target.value) })
10831
11141
  }
@@ -11570,7 +11880,18 @@ function RecognitionEngineSettings({
11570
11880
  settings,
11571
11881
  updateSettings
11572
11882
  }
11573
- ), engine === "whisper-http" && /* @__PURE__ */ import_react46.default.createElement(SettingsSubcard, { title: commons.connection.title(), icon: "link" }, /* @__PURE__ */ import_react46.default.createElement("p", null, recognition.whisper.captureHelp()), /* @__PURE__ */ import_react46.default.createElement(WhisperSettings, { controller })), engine === "qwen-http" && /* @__PURE__ */ import_react46.default.createElement(SettingsSubcard, { title: commons.connection.title(), icon: "link" }, /* @__PURE__ */ import_react46.default.createElement("p", null, recognition.qwen.captureHelp()), settings.engine !== "qwen-http" && /* @__PURE__ */ import_react46.default.createElement(QwenSettings, { controller })));
11883
+ ), ["whisper-http", "qwen-http"].includes(engine) && /* @__PURE__ */ import_react46.default.createElement(SettingsSubcard, { title: commons.connection.title(), icon: "link" }, engine === "whisper-http" ? /* @__PURE__ */ import_react46.default.createElement(WhisperSettings, { controller }) : settings.engine !== "qwen-http" && /* @__PURE__ */ import_react46.default.createElement(QwenSettings, { controller }), /* @__PURE__ */ import_react46.default.createElement(
11884
+ TextAreaField,
11885
+ {
11886
+ label: recognition.context.label(),
11887
+ hint: recognition.context.hint(),
11888
+ placeholder: recognition.context.placeholder(),
11889
+ value: settings.recognitionContext || "",
11890
+ rows: 3,
11891
+ maxLength: TRANSCRIPTION_CONTEXT_MAX_LENGTH,
11892
+ onCommit: (recognitionContext) => updateSettings({ recognitionContext })
11893
+ }
11894
+ )));
11574
11895
  }
11575
11896
 
11576
11897
  // src/modules/settings/sections/recognition/RecognitionFilterSettings.tsx
@@ -13324,7 +13645,7 @@ async function removeBrowserModel(key, storage = caches) {
13324
13645
  }
13325
13646
 
13326
13647
  // src/modules/recognition/local/assets.ts
13327
- var browserAsrAssetBase = () => "/api/dsh-live-voice/asr/" + encodeURIComponent(true ? "0.6.1" : "test") + "/";
13648
+ var browserAsrAssetBase = () => "/api/dsh-live-voice/asr/" + encodeURIComponent(true ? "0.6.2" : "test") + "/";
13328
13649
 
13329
13650
  // src/modules/recognition/local/runtimeIdentity.ts
13330
13651
  var workerProtocol = "asr-worker-v3-ort-1.31.0-dev.20260914-8d85527a0";
@@ -13792,7 +14113,7 @@ var BrowserLocalRecognitionEngine = class extends WhisperHttpRecognitionEngine {
13792
14113
  message;
13793
14114
  release;
13794
14115
  constructor({ manager, modelKey, device = "webgpu", message: message2, ...options }) {
13795
- const margins = ((options.recordingPreRollMs ?? 320) + (options.recordingPostRollMs ?? 320)) / 1e3;
14116
+ const margins = (Math.max(TRANSCRIPTION_BOUNDARY_MARGIN_MS, options.recordingPreRollMs ?? 320) + Math.max(TRANSCRIPTION_BOUNDARY_MARGIN_MS, options.recordingPostRollMs ?? 320)) / 1e3;
13796
14117
  super({
13797
14118
  ...options,
13798
14119
  maxUtteranceSeconds: Math.max(
@@ -14034,6 +14355,15 @@ function apply(ctx) {
14034
14355
  const settingsControllers = /* @__PURE__ */ new Set();
14035
14356
  let disposed = false;
14036
14357
  const recordings = createRecordingRepository();
14358
+ const recordingTimeline = () => {
14359
+ const append = recordings.timeline();
14360
+ return (kind) => {
14361
+ void append(kind).catch(() => {
14362
+ for (const entry of controllers.values())
14363
+ entry.controller.patch({ error: t("dsh-live-voice.recording.error.failed") });
14364
+ });
14365
+ };
14366
+ };
14037
14367
  const localModels = new LocalModelManager();
14038
14368
  void localModels.refresh();
14039
14369
  const leaveGuard = createTranscriptionLeaveGuard(window);
@@ -14196,6 +14526,7 @@ function apply(ctx) {
14196
14526
  "browserQwenModel",
14197
14527
  "browserAsrDevice",
14198
14528
  "recognitionLang",
14529
+ "recognitionContext",
14199
14530
  "inputDeviceId",
14200
14531
  "recognitionProcessLocally",
14201
14532
  "recognitionAutoInstall",
@@ -14215,6 +14546,7 @@ function apply(ctx) {
14215
14546
  const changesRecognition = (next) => Object.keys(next).some((key) => recognitionSettingKeys.has(key));
14216
14547
  const recordingOptions = (settings, inputSource) => ({
14217
14548
  recordings,
14549
+ recognitionContext: settings.recognitionContext,
14218
14550
  inputSource,
14219
14551
  sourceId: inputSource,
14220
14552
  sourceLabel: inputSource === "shared" ? settings.sharedSourceName : settings.micSourceName,
@@ -14415,6 +14747,7 @@ function apply(ctx) {
14415
14747
  const meter = new MicrophoneMeter();
14416
14748
  const recognition = recognitionFor(settings, meter);
14417
14749
  entry.controller = new VoiceCoordinator({
14750
+ recordingTimeline,
14418
14751
  persistSettings: (next) => savePreferences(entry.controller, next),
14419
14752
  recognition,
14420
14753
  engines: { browser: engineBrowser, say: engineSay, "qwen-http": engineQwen },
@@ -14818,6 +15151,7 @@ function apply(ctx) {
14818
15151
  const meter = new MicrophoneMeter();
14819
15152
  const recognition = recognitionFor(settings, meter);
14820
15153
  const c = new VoiceCoordinator({
15154
+ recordingTimeline,
14821
15155
  persistSettings: (next) => savePreferences(c, next),
14822
15156
  recognition,
14823
15157
  engines: {