dsh-live-voice 0.6.0 → 0.6.2

package/lib/client.js

@@ -152,7 +152,7 @@ var SessionSpeechMemory = class {
 };
 
 // src/modules/recognition/vad/assets.ts
-var sileroVadAssetBase = () => "/api/dsh-live-voice/vad/" + encodeURIComponent(true ? "0.6.0" : "test") + "/";
+var sileroVadAssetBase = () => "/api/dsh-live-voice/vad/" + encodeURIComponent(true ? "0.6.2" : "test") + "/";
 async function sileroVadAvailable(fetchImpl = globalThis.fetch) {
   try {
     const response = await fetchImpl(sileroVadAssetBase() + "availability", {
@@ -615,6 +615,7 @@ var RecordingRepositoryError = class extends Error {
   }
   code;
 };
+var audioPageBytes = (page) => (page.decoderConfig?.description?.byteLength ?? 0) + page.packets.reduce((total, packet) => total + packet.data.byteLength, 0);
 var globals = () => globalThis;
 function unsupported(cause) {
   return new RecordingRepositoryError(
@@ -919,6 +920,19 @@ async function openRecordingStore() {
         }
       };
     }),
+    audioBytes: (id) => transact("readonly", (tx, result) => {
+      let bytes = 0;
+      const request = tx.objectStore("pages").index("recordingId").openCursor(IDBKeyRange.only(id));
+      request.onsuccess = () => {
+        const cursor = request.result;
+        if (!cursor) {
+          result(bytes);
+          return;
+        }
+        bytes += audioPageBytes(cursor.value);
+        cursor.continue();
+      };
+    }),
     nextPage: (id, start, end) => transact("readonly", (tx, result) => {
       if (end <= start) {
         result(null);
@@ -956,6 +970,7 @@ function createMemoryStore(maxBytes = 32 * 1024 * 1024) {
   let used = 0;
   const size2 = (page) => 256 + (page.decoderConfig?.description?.byteLength ?? 0) + page.packets.reduce((total, packet) => total + packet.data.byteLength + 64, 0);
   return {
+    audioBytes: async (id) => (pages.get(id) ?? []).reduce((sum, page) => sum + audioPageBytes(page), 0),
     list: async () => structuredClone([...sessions.values()]),
     get: async (id) => structuredClone(sessions.get(id) ?? null),
     create: async (session) => {
@@ -1204,6 +1219,17 @@ function createRecordingRepository() {
     }
     return locate(id);
   }
+  const measured = /* @__PURE__ */ new Map();
+  async function withAudioBytes(session) {
+    if (session.audioBytes !== void 0) return session;
+    let cached = measured.get(session.id);
+    if (!cached || cached.samples !== session.savedSamples) {
+      const { store } = await locate(session.id);
+      cached = { samples: session.savedSamples, bytes: await store.audioBytes(session.id) };
+      measured.set(session.id, cached);
+    }
+    return { ...session, audioBytes: cached.bytes };
+  }
   const repository = {
     version: 1,
     segments: recordingSegments,
@@ -1214,14 +1240,21 @@ function createRecordingRepository() {
     },
     async list() {
       check2();
-      return (await all()).map(
-        (session) => writers.get(session.id)?.writer.session ?? cloneSession(session)
+      const sessions = await all();
+      const ids = new Set(sessions.map((session) => session.id));
+      for (const id of measured.keys()) if (!ids.has(id)) measured.delete(id);
+      return Promise.all(
+        sessions.map(
+          (session) => withAudioBytes(writers.get(session.id)?.writer.session ?? cloneSession(session))
+        )
       );
     },
     async get(id) {
       check2();
       try {
-        return writers.get(id)?.writer.session ?? cloneSession((await locate(id)).session);
+        return await withAudioBytes(
+          writers.get(id)?.writer.session ?? cloneSession((await locate(id)).session)
+        );
       } catch (error) {
         if (error instanceof RecordingRepositoryError && error.code === "not-found") return null;
         throw error;
@@ -1326,6 +1359,66 @@ function createRecordingRepository() {
         emit();
       });
     },
+    async setName(id, value) {
+      check2();
+      if (typeof value !== "string" || value.trim().length > 120 || /[\u0000-\u001f\u007f-\u009f\u2028\u2029]/u.test(value))
+        throw new RecordingRepositoryError(
+          "invalid-range",
+          "Recording names must be a single line of at most 120 characters."
+        );
+      const name = value.trim();
+      const active = writers.get(id);
+      if (active) return active.setName(name);
+      await serial(async () => {
+        const { store } = await locate(id);
+        const release = store === memory ? () => {
+        } : await holdLock(id);
+        try {
+          await store.update(id, (current) => {
+            if (!current.starred)
+              throw new RecordingRepositoryError(
+                "invalid-range",
+                "Star a recording before naming it."
+              );
+            return { ...current, name };
+          });
+        } finally {
+          release();
+        }
+        emit();
+      });
+    },
+    timeline() {
+      const targets = [...writers.entries()];
+      return async (kind) => {
+        const occurredAt = now();
+        await Promise.all(
+          targets.map(async ([id, active]) => {
+            const value = {
+              id: crypto.randomUUID(),
+              kind,
+              occurredAt,
+              atSample: active.writer.session.capturedSamples
+            };
+            if (writers.get(id) === active) return active.event(value);
+            await serial(async () => {
+              const { store } = await locate(id);
+              const release = store === memory ? () => {
+              } : await holdLock(id);
+              try {
+                await store.update(id, (current) => ({
+                  ...current,
+                  events: [...current.events ?? [], value]
+                }));
+              } finally {
+                release();
+              }
+              emit();
+            });
+          })
+        );
+      };
+    },
     async begin(input) {
       check2();
       return serial(async () => {
@@ -1363,12 +1456,16 @@ function createRecordingRepository() {
           state: "recording",
           capturedSamples: 0,
           savedSamples: 0,
+          audioBytes: 0,
           persistent,
           codec: "opus",
           storage: policy.enabled ? "indexeddb" : "memory",
           configuration: input.configuration,
           pins: [],
           attempts: [],
+          events: [
+            { id: crypto.randomUUID(), kind: "capture-start", occurredAt: now(), atSample: 0 }
+          ],
           keepPrevious: Math.max(1, Math.floor(policy.keepPrevious))
         };
         try {
@@ -1422,7 +1519,14 @@ function createRecordingRepository() {
             sampleRate,
             ...encoded
           };
-          await update((current) => ({ ...current, savedSamples: page.startSample + count }), page);
+          await update(
+            (current) => ({
+              ...current,
+              savedSamples: page.startSample + count,
+              audioBytes: (current.audioBytes ?? 0) + audioPageBytes(page)
+            }),
+            page
+          );
           pendingCount = 0;
           index++;
         };
@@ -1553,7 +1657,16 @@ function createRecordingRepository() {
                   ...current,
                   state: fatal ? "failed" : state,
                   error: fatal?.code ?? error,
-                  endedAt: now()
+                  endedAt: now(),
+                  events: [
+                    ...current.events ?? [],
+                    {
+                      id: crypto.randomUUID(),
+                      kind: "capture-end",
+                      occurredAt: now(),
+                      atSample: current.capturedSamples
+                    }
+                  ]
                 }));
               } finally {
                 pending = new Float32Array(0);
@@ -1572,6 +1685,25 @@ function createRecordingRepository() {
           setStarred: (value) => enqueue(async () => {
             await update((current) => ({ ...current, starred: value }));
           }),
+          setName: (name) => enqueue(async () => {
+            await update((current) => {
+              if (!current.starred)
+                throw new RecordingRepositoryError(
+                  "invalid-range",
+                  "Star a recording before naming it."
+                );
+              return { ...current, name };
+            });
+          }),
+          event: (value) => enqueue(async () => {
+            await update((current) => ({
+              ...current,
+              events: [
+                ...current.events ?? [],
+                { ...value, atSample: current.capturedSamples }
+              ]
+            }));
+          }),
           idle: async () => {
             await tail;
           }
@@ -1734,8 +1866,10 @@ var en = {
   "dsh-live-voice.commons.update.cancel": "Cancel",
   "dsh-live-voice.commons.update.confirm": "Install Live Voice {version} in this DSH profile? Download requires network access. Restart DSH afterward when requested; nothing restarts automatically.",
   "dsh-live-voice.commons.update.current": "This version is already installed. Restart DSH to use it.",
+  "dsh-live-voice.commons.update.delayed": "The requested version is not available to this profile yet. It may still be publishing or be delayed by pnpm minimumReleaseAge. Wait, then check the DSH plugin manager; no security policy was bypassed.",
   "dsh-live-voice.commons.update.done": "Live Voice {version} installed. Reload the page when ready.",
   "dsh-live-voice.commons.update.failed": "Update failed. The release may not be published on npm yet. Check the DSH plugin manager.",
+  "dsh-live-voice.commons.update.incompatible": "This release is incompatible with the running DSH version. Check the DSH plugin manager and use a compatible DSH/plugin version; no compatibility exemption was granted.",
   "dsh-live-voice.commons.update.installing": "Updating\u2026",
   "dsh-live-voice.commons.update.label": "Update available",
   "dsh-live-voice.commons.update.link": "Update available: {version}. Open release",
@@ -1805,6 +1939,9 @@ var en = {
   "dsh-live-voice.recognition.commands.send": "Send to running agent",
   "dsh-live-voice.recognition.commands.stopSpeech": "Stop assistant speech",
   "dsh-live-voice.recognition.commands.title": "Voice commands",
+  "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.",
+  "dsh-live-voice.recognition.context.label": "Transcription context",
+  "dsh-live-voice.recognition.context.placeholder": "Native Brazilian Portuguese speaker who frequently uses technical terms in English.",
   "dsh-live-voice.recognition.dictation.cancel": "Cancel dictation",
   "dsh-live-voice.recognition.engine.label": "Recognition engine",
   "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.",
@@ -2077,7 +2214,7 @@ var en = {
   "dsh-live-voice.speak.rate.label": "Speech rate",
   "dsh-live-voice.speak.responseDelay.help": "After you stop speaking, automatic assistant playback waits for this much continuous silence. Speaking again restarts the wait.",
   "dsh-live-voice.speak.responseDelay.label": "Assistant response delay",
-  "dsh-live-voice.speak.responseDelay.none": "No delay",
+  "dsh-live-voice.speak.responseDelay.oneAndHalf": "1.5 seconds",
   "dsh-live-voice.speak.segmentGap.help": "Wait this many milliseconds between consecutive spoken segments. 400 ms is the default.",
   "dsh-live-voice.speak.segmentGap.label": "Pause between speech segments",
   "dsh-live-voice.speak.status.paused": "Speech paused",
@@ -2129,8 +2266,10 @@ var es = {
   "dsh-live-voice.commons.update.cancel": "Cancelar",
   "dsh-live-voice.commons.update.confirm": "\xBFInstalar Voz en vivo {version} en este perfil de DSH? La descarga requiere red. Reinicia DSH despu\xE9s si se solicita; nada se reinicia autom\xE1ticamente.",
   "dsh-live-voice.commons.update.current": "Esta versi\xF3n ya est\xE1 instalada. Reinicia DSH para usarla.",
+  "dsh-live-voice.commons.update.delayed": "La versi\xF3n solicitada a\xFAn no est\xE1 disponible para este perfil. Puede estar public\xE1ndose o esperando el minimumReleaseAge de pnpm. Espera y consulta el gestor de plugins de DSH; no se omiti\xF3 ninguna pol\xEDtica de seguridad.",
   "dsh-live-voice.commons.update.done": "Voz en vivo {version} instalado. Recarga la p\xE1gina cuando quieras.",
   "dsh-live-voice.commons.update.failed": "La actualizaci\xF3n fall\xF3. Puede que la versi\xF3n a\xFAn no est\xE9 en npm. Revisa el gestor de plugins de DSH.",
+  "dsh-live-voice.commons.update.incompatible": "Esta versi\xF3n es incompatible con el DSH en ejecuci\xF3n. Consulta el gestor de plugins y usa versiones compatibles de DSH y del plugin; no se concedi\xF3 ninguna excepci\xF3n.",
   "dsh-live-voice.commons.update.installing": "Actualizando\u2026",
   "dsh-live-voice.commons.update.label": "Actualizaci\xF3n disponible",
   "dsh-live-voice.commons.update.link": "Actualizaci\xF3n disponible: {version}. Abrir la p\xE1gina de esta versi\xF3n",
@@ -2200,6 +2339,9 @@ var es = {
   "dsh-live-voice.recognition.commands.send": "Enviar al agente en ejecuci\xF3n",
   "dsh-live-voice.recognition.commands.stopSpeech": "Detener la lectura en voz alta del asistente",
   "dsh-live-voice.recognition.commands.title": "Comandos de voz",
+  "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.",
+  "dsh-live-voice.recognition.context.label": "Contexto de transcripci\xF3n",
+  "dsh-live-voice.recognition.context.placeholder": "Native Brazilian Portuguese speaker who frequently uses technical terms in English.",
   "dsh-live-voice.recognition.dictation.cancel": "Cancelar dictado",
   "dsh-live-voice.recognition.engine.label": "Motor de reconocimiento",
   "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.",
@@ -2472,7 +2614,7 @@ var es = {
   "dsh-live-voice.speak.rate.label": "Velocidad de habla",
   "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.",
   "dsh-live-voice.speak.responseDelay.label": "Demora de respuesta del asistente",
-  "dsh-live-voice.speak.responseDelay.none": "Sin retraso",
+  "dsh-live-voice.speak.responseDelay.oneAndHalf": "1,5 segundos",
   "dsh-live-voice.speak.segmentGap.help": "Espera esta cantidad de milisegundos entre segmentos hablados consecutivos. El valor predeterminado es 400 ms.",
   "dsh-live-voice.speak.segmentGap.label": "Pausa entre segmentos de voz",
   "dsh-live-voice.speak.status.paused": "Lectura en pausa",
@@ -2524,8 +2666,10 @@ var fr = {
   "dsh-live-voice.commons.update.cancel": "Annuler",
   "dsh-live-voice.commons.update.confirm": "Installer Voix en direct {version} dans ce profil DSH ? Le t\xE9l\xE9chargement n\xE9cessite le r\xE9seau. Red\xE9marrez DSH ensuite si demand\xE9 ; rien ne red\xE9marre automatiquement.",
   "dsh-live-voice.commons.update.current": "Cette version est d\xE9j\xE0 install\xE9e. Red\xE9marrez DSH pour l\u2019utiliser.",
+  "dsh-live-voice.commons.update.delayed": "La version demand\xE9e n\u2019est pas encore disponible pour ce profil. La publication ou le minimumReleaseAge de pnpm peut la retarder. Patientez puis consultez le gestionnaire DSH ; aucune r\xE8gle de s\xE9curit\xE9 n\u2019a \xE9t\xE9 contourn\xE9e.",
   "dsh-live-voice.commons.update.done": "Voix en direct {version} install\xE9. Rechargez la page quand vous le souhaitez.",
   "dsh-live-voice.commons.update.failed": "\xC9chec de mise \xE0 jour. La version n\u2019est peut-\xEAtre pas encore sur npm. Consultez le gestionnaire de plugins DSH.",
+  "dsh-live-voice.commons.update.incompatible": "Cette version est incompatible avec le DSH en cours. Consultez le gestionnaire de plugins et utilisez des versions compatibles de DSH et du plugin ; aucune d\xE9rogation n\u2019a \xE9t\xE9 accord\xE9e.",
   "dsh-live-voice.commons.update.installing": "Mise \xE0 jour\u2026",
   "dsh-live-voice.commons.update.label": "Mise \xE0 jour disponible",
   "dsh-live-voice.commons.update.link": "Mise \xE0 jour disponible : {version}. Ouvrir la page de cette version",
@@ -2595,6 +2739,9 @@ var fr = {
   "dsh-live-voice.recognition.commands.send": "Envoyer \xE0 l\u2019agent en cours d\u2019ex\xE9cution",
   "dsh-live-voice.recognition.commands.stopSpeech": "Arr\xEAter la lecture vocale de l\u2019assistant",
   "dsh-live-voice.recognition.commands.title": "Commandes vocales",
+  "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.",
+  "dsh-live-voice.recognition.context.label": "Contexte de transcription",
+  "dsh-live-voice.recognition.context.placeholder": "Native Brazilian Portuguese speaker who frequently uses technical terms in English.",
   "dsh-live-voice.recognition.dictation.cancel": "Annuler la dict\xE9e",
   "dsh-live-voice.recognition.engine.label": "Moteur de reconnaissance",
   "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.",
@@ -2867,7 +3014,7 @@ var fr = {
   "dsh-live-voice.speak.rate.label": "D\xE9bit de parole",
   "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.",
   "dsh-live-voice.speak.responseDelay.label": "D\xE9lai de r\xE9ponse de l\u2019assistant",
-  "dsh-live-voice.speak.responseDelay.none": "Sans d\xE9lai",
+  "dsh-live-voice.speak.responseDelay.oneAndHalf": "1,5 seconde",
   "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.",
   "dsh-live-voice.speak.segmentGap.label": "Pause entre les segments vocaux",
   "dsh-live-voice.speak.status.paused": "Lecture en pause",
@@ -2919,8 +3066,10 @@ var hi = {
   "dsh-live-voice.commons.update.cancel": "\u0930\u0926\u094D\u0926 \u0915\u0930\u0947\u0902",
   "dsh-live-voice.commons.update.confirm": "\u0907\u0938 DSH \u092A\u094D\u0930\u094B\u092B\u093C\u093E\u0907\u0932 \u092E\u0947\u0902 \u0932\u093E\u0907\u0935 \u0935\u0949\u0907\u0938 {version} \u0907\u0902\u0938\u094D\u091F\u0949\u0932 \u0915\u0930\u0947\u0902? \u0921\u093E\u0909\u0928\u0932\u094B\u0921 \u0915\u0947 \u0932\u093F\u090F \u0928\u0947\u091F\u0935\u0930\u094D\u0915 \u091A\u093E\u0939\u093F\u090F\u0964 \u0905\u0928\u0941\u0930\u094B\u0927 \u0939\u094B\u0928\u0947 \u092A\u0930 \u092C\u093E\u0926 \u092E\u0947\u0902 DSH \u092A\u0941\u0928\u0903 \u0936\u0941\u0930\u0942 \u0915\u0930\u0947\u0902; \u0915\u0941\u091B \u092D\u0940 \u0905\u092A\u0928\u0947 \u0906\u092A \u092A\u0941\u0928\u0903 \u0936\u0941\u0930\u0942 \u0928\u0939\u0940\u0902 \u0939\u094B\u0917\u093E\u0964",
   "dsh-live-voice.commons.update.current": "\u092F\u0939 \u0938\u0902\u0938\u094D\u0915\u0930\u0923 \u092A\u0939\u0932\u0947 \u0938\u0947 \u0907\u0902\u0938\u094D\u091F\u0949\u0932 \u0939\u0948\u0964 \u0909\u092A\u092F\u094B\u0917 \u0915\u0947 \u0932\u093F\u090F DSH \u092A\u0941\u0928\u0903 \u0936\u0941\u0930\u0942 \u0915\u0930\u0947\u0902\u0964",
+  "dsh-live-voice.commons.update.delayed": "\u092E\u093E\u0901\u0917\u093E \u0917\u092F\u093E \u0938\u0902\u0938\u094D\u0915\u0930\u0923 \u0905\u092D\u0940 \u0907\u0938 \u092A\u094D\u0930\u094B\u092B\u093C\u093E\u0907\u0932 \u0915\u0947 \u0932\u093F\u090F \u0909\u092A\u0932\u092C\u094D\u0927 \u0928\u0939\u0940\u0902 \u0939\u0948\u0964 \u092A\u094D\u0930\u0915\u093E\u0936\u0928 \u091C\u093E\u0930\u0940 \u0939\u094B \u0938\u0915\u0924\u093E \u0939\u0948 \u092F\u093E pnpm \u0915\u093E minimumReleaseAge \u0935\u093F\u0932\u0902\u092C \u0915\u0930 \u0938\u0915\u0924\u093E \u0939\u0948\u0964 \u092A\u094D\u0930\u0924\u0940\u0915\u094D\u0937\u093E \u0915\u0930\u0947\u0902, \u092B\u093F\u0930 DSH \u092A\u094D\u0932\u0917\u0907\u0928 \u092A\u094D\u0930\u092C\u0902\u0927\u0915 \u0926\u0947\u0916\u0947\u0902; \u0915\u094B\u0908 \u0938\u0941\u0930\u0915\u094D\u0937\u093E \u0928\u0940\u0924\u093F \u0928\u0939\u0940\u0902 \u092C\u0926\u0932\u0940 \u0917\u0908\u0964",
   "dsh-live-voice.commons.update.done": "\u0932\u093E\u0907\u0935 \u0935\u0949\u0907\u0938 {version} \u0907\u0902\u0938\u094D\u091F\u0949\u0932 \u0939\u094B \u0917\u092F\u093E\u0964 \u0924\u0948\u092F\u093E\u0930 \u0939\u094B\u0928\u0947 \u092A\u0930 \u092A\u0947\u091C \u0930\u0940\u0932\u094B\u0921 \u0915\u0930\u0947\u0902\u0964",
   "dsh-live-voice.commons.update.failed": "\u0905\u092A\u0921\u0947\u091F \u0935\u093F\u092B\u0932\u0964 \u0930\u093F\u0932\u0940\u091C\u093C \u0905\u092D\u0940 npm \u092A\u0930 \u0909\u092A\u0932\u092C\u094D\u0927 \u0928\u0939\u0940\u0902 \u0939\u094B \u0938\u0915\u0924\u0940\u0964 DSH \u092A\u094D\u0932\u0917\u0907\u0928 \u092A\u094D\u0930\u092C\u0902\u0927\u0915 \u0926\u0947\u0916\u0947\u0902\u0964",
+  "dsh-live-voice.commons.update.incompatible": "\u092F\u0939 \u0930\u093F\u0932\u0940\u091C\u093C \u091A\u0932 \u0930\u0939\u0947 DSH \u0938\u0902\u0938\u094D\u0915\u0930\u0923 \u0915\u0947 \u0938\u093E\u0925 \u0938\u0902\u0917\u0924 \u0928\u0939\u0940\u0902 \u0939\u0948\u0964 DSH \u092A\u094D\u0932\u0917\u0907\u0928 \u092A\u094D\u0930\u092C\u0902\u0927\u0915 \u0926\u0947\u0916\u0947\u0902 \u0914\u0930 \u0938\u0902\u0917\u0924 DSH \u0924\u0925\u093E \u092A\u094D\u0932\u0917\u0907\u0928 \u0938\u0902\u0938\u094D\u0915\u0930\u0923 \u0909\u092A\u092F\u094B\u0917 \u0915\u0930\u0947\u0902; \u0915\u094B\u0908 \u0938\u0902\u0917\u0924\u0924\u093E \u091B\u0942\u091F \u0928\u0939\u0940\u0902 \u0926\u0940 \u0917\u0908\u0964",
   "dsh-live-voice.commons.update.installing": "\u0905\u092A\u0921\u0947\u091F \u0939\u094B \u0930\u0939\u093E \u0939\u0948\u2026",
   "dsh-live-voice.commons.update.label": "\u0905\u092A\u0921\u0947\u091F \u0909\u092A\u0932\u092C\u094D\u0927 \u0939\u0948",
   "dsh-live-voice.commons.update.link": "\u0905\u092A\u0921\u0947\u091F \u0909\u092A\u0932\u092C\u094D\u0927 \u0939\u0948: {version}\u0964 \u0930\u093F\u0932\u0940\u091C\u093C \u0916\u094B\u0932\u0947\u0902",
@@ -2990,6 +3139,9 @@ var hi = {
   "dsh-live-voice.recognition.commands.send": "\u091A\u0932 \u0930\u0939\u0947 \u090F\u091C\u0947\u0902\u091F \u0915\u094B \u092D\u0947\u091C\u0947\u0902",
   "dsh-live-voice.recognition.commands.stopSpeech": "\u0938\u0939\u093E\u092F\u0915 \u0915\u093E \u0935\u093E\u091A\u0928 \u092C\u0902\u0926 \u0915\u0930\u0947\u0902",
   "dsh-live-voice.recognition.commands.title": "\u0935\u0949\u0907\u0938 \u0915\u092E\u093E\u0902\u0921",
+  "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",
+  "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",
+  "dsh-live-voice.recognition.context.placeholder": "Native Brazilian Portuguese speaker who frequently uses technical terms in English.",
   "dsh-live-voice.recognition.dictation.cancel": "\u0921\u093F\u0915\u094D\u091F\u0947\u0936\u0928 \u0930\u0926\u094D\u0926 \u0915\u0930\u0947\u0902",
   "dsh-live-voice.recognition.engine.label": "\u0935\u093E\u0915\u094D \u092A\u0939\u091A\u093E\u0928 \u0907\u0902\u091C\u0928",
   "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",
@@ -3262,7 +3414,7 @@ var hi = {
   "dsh-live-voice.speak.rate.label": "\u092C\u094B\u0932\u0928\u0947 \u0915\u0940 \u0917\u0924\u093F",
   "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",
   "dsh-live-voice.speak.responseDelay.label": "\u0938\u0939\u093E\u092F\u0915 \u0915\u0947 \u091C\u0935\u093E\u092C \u0915\u093E \u0935\u093F\u0932\u0902\u092C",
-  "dsh-live-voice.speak.responseDelay.none": "\u0915\u094B\u0908 \u0935\u093F\u0932\u0902\u092C \u0928\u0939\u0940\u0902",
+  "dsh-live-voice.speak.responseDelay.oneAndHalf": "1.5 \u0938\u0947\u0915\u0902\u0921",
   "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",
   "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",
   "dsh-live-voice.speak.status.paused": "\u0935\u093E\u091A\u0928 \u0920\u0939\u0930\u093E \u0939\u0941\u0906 \u0939\u0948",
@@ -3314,8 +3466,10 @@ var ptBR = {
   "dsh-live-voice.commons.update.cancel": "Cancelar",
   "dsh-live-voice.commons.update.confirm": "Instalar Voz ao Vivo {version} neste perfil do DSH? O download precisa de rede. Reinicie o DSH depois, se solicitado; nada reinicia automaticamente.",
   "dsh-live-voice.commons.update.current": "Esta vers\xE3o j\xE1 est\xE1 instalada. Reinicie o DSH para utiliz\xE1-la.",
+  "dsh-live-voice.commons.update.delayed": "A vers\xE3o solicitada ainda n\xE3o est\xE1 dispon\xEDvel para este perfil. A publica\xE7\xE3o pode estar em andamento ou aguardando o minimumReleaseAge do pnpm. Aguarde e consulte o gerenciador de plugins do DSH; nenhuma pol\xEDtica de seguran\xE7a foi ignorada.",
   "dsh-live-voice.commons.update.done": "Voz ao Vivo {version} instalado. Recarregue a p\xE1gina quando quiser.",
   "dsh-live-voice.commons.update.failed": "Falha na atualiza\xE7\xE3o. A vers\xE3o pode ainda n\xE3o estar no npm. Verifique o gerenciador de plugins do DSH.",
+  "dsh-live-voice.commons.update.incompatible": "Esta vers\xE3o \xE9 incompat\xEDvel com o DSH em execu\xE7\xE3o. Consulte o gerenciador de plugins e use vers\xF5es compat\xEDveis do DSH e do plugin; nenhuma exce\xE7\xE3o de compatibilidade foi concedida.",
   "dsh-live-voice.commons.update.installing": "Atualizando\u2026",
   "dsh-live-voice.commons.update.label": "Atualiza\xE7\xE3o dispon\xEDvel",
   "dsh-live-voice.commons.update.link": "Atualiza\xE7\xE3o dispon\xEDvel: {version}. Abrir lan\xE7amento",
@@ -3385,6 +3539,9 @@ var ptBR = {
   "dsh-live-voice.recognition.commands.send": "Enviar ao agente em execu\xE7\xE3o",
   "dsh-live-voice.recognition.commands.stopSpeech": "Parar fala do assistente",
   "dsh-live-voice.recognition.commands.title": "Comandos de voz",
+  "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.",
+  "dsh-live-voice.recognition.context.label": "Contexto da transcri\xE7\xE3o",
+  "dsh-live-voice.recognition.context.placeholder": "Native Brazilian Portuguese speaker who frequently uses technical terms in English.",
   "dsh-live-voice.recognition.dictation.cancel": "Cancelar ditado",
   "dsh-live-voice.recognition.engine.label": "Mecanismo de reconhecimento",
   "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.",
@@ -3657,7 +3814,7 @@ var ptBR = {
   "dsh-live-voice.speak.rate.label": "Velocidade da fala",
   "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.",
   "dsh-live-voice.speak.responseDelay.label": "Atraso da resposta do assistente",
-  "dsh-live-voice.speak.responseDelay.none": "Sem atraso",
+  "dsh-live-voice.speak.responseDelay.oneAndHalf": "1,5 segundo",
   "dsh-live-voice.speak.segmentGap.help": "Aguarde esta quantidade de milissegundos entre trechos falados consecutivos. O padr\xE3o \xE9 400 ms.",
   "dsh-live-voice.speak.segmentGap.label": "Pausa entre trechos de fala",
   "dsh-live-voice.speak.status.paused": "Fala pausada",
@@ -3709,8 +3866,10 @@ var zh = {
   "dsh-live-voice.commons.update.cancel": "\u53D6\u6D88",
   "dsh-live-voice.commons.update.confirm": "\u5728\u6B64 DSH \u914D\u7F6E\u4E2D\u5B89\u88C5 \u5B9E\u65F6\u8BED\u97F3 {version}\uFF1F\u4E0B\u8F7D\u9700\u8981\u7F51\u7EDC\u3002\u82E5\u63D0\u793A\uFF0C\u8BF7\u4E4B\u540E\u91CD\u542F DSH\uFF1B\u4E0D\u4F1A\u81EA\u52A8\u91CD\u542F\u3002",
   "dsh-live-voice.commons.update.current": "\u6B64\u7248\u672C\u5DF2\u5B89\u88C5\u3002\u8BF7\u91CD\u542F DSH \u540E\u4F7F\u7528\u3002",
+  "dsh-live-voice.commons.update.delayed": "\u6B64\u914D\u7F6E\u6682\u65F6\u65E0\u6CD5\u83B7\u53D6\u6240\u8BF7\u6C42\u7684\u7248\u672C\u3002\u5B83\u53EF\u80FD\u4ECD\u5728\u53D1\u5E03\uFF0C\u6216\u53D7 pnpm minimumReleaseAge \u9650\u5236\u3002\u8BF7\u7A0D\u540E\u68C0\u67E5 DSH \u63D2\u4EF6\u7BA1\u7406\u5668\uFF1B\u672A\u7ED5\u8FC7\u4EFB\u4F55\u5B89\u5168\u7B56\u7565\u3002",
   "dsh-live-voice.commons.update.done": "\u5B9E\u65F6\u8BED\u97F3 {version} \u5DF2\u5B89\u88C5\u3002\u51C6\u5907\u597D\u540E\u8BF7\u5237\u65B0\u9875\u9762\u3002",
   "dsh-live-voice.commons.update.failed": "\u66F4\u65B0\u5931\u8D25\u3002\u8BE5\u7248\u672C\u53EF\u80FD\u5C1A\u672A\u53D1\u5E03\u5230 npm\u3002\u8BF7\u67E5\u770B DSH \u63D2\u4EF6\u7BA1\u7406\u5668\u3002",
+  "dsh-live-voice.commons.update.incompatible": "\u6B64\u7248\u672C\u4E0E\u6B63\u5728\u8FD0\u884C\u7684 DSH \u4E0D\u517C\u5BB9\u3002\u8BF7\u68C0\u67E5 DSH \u63D2\u4EF6\u7BA1\u7406\u5668\u5E76\u4F7F\u7528\u517C\u5BB9\u7684 DSH \u548C\u63D2\u4EF6\u7248\u672C\uFF1B\u672A\u6388\u4E88\u517C\u5BB9\u6027\u8C41\u514D\u3002",
   "dsh-live-voice.commons.update.installing": "\u6B63\u5728\u66F4\u65B0\u2026",
   "dsh-live-voice.commons.update.label": "\u6709\u53EF\u7528\u66F4\u65B0",
   "dsh-live-voice.commons.update.link": "\u6709\u53EF\u7528\u66F4\u65B0\uFF1A{version}\u3002\u6253\u5F00\u53D1\u5E03\u9875\u9762",
@@ -3780,6 +3939,9 @@ var zh = {
   "dsh-live-voice.recognition.commands.send": "\u53D1\u9001\u7ED9\u8FD0\u884C\u4E2D\u7684\u667A\u80FD\u4F53",
   "dsh-live-voice.recognition.commands.stopSpeech": "\u505C\u6B62\u52A9\u624B\u6717\u8BFB",
   "dsh-live-voice.recognition.commands.title": "\u8BED\u97F3\u547D\u4EE4",
+  "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",
+  "dsh-live-voice.recognition.context.label": "\u8F6C\u5199\u4E0A\u4E0B\u6587",
+  "dsh-live-voice.recognition.context.placeholder": "Native Brazilian Portuguese speaker who frequently uses technical terms in English.",
   "dsh-live-voice.recognition.dictation.cancel": "\u53D6\u6D88\u542C\u5199",
   "dsh-live-voice.recognition.engine.label": "\u8BC6\u522B\u5F15\u64CE",
   "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",
@@ -4052,7 +4214,7 @@ var zh = {
   "dsh-live-voice.speak.rate.label": "\u8BED\u901F",
   "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",
   "dsh-live-voice.speak.responseDelay.label": "\u52A9\u624B\u54CD\u5E94\u5EF6\u8FDF",
-  "dsh-live-voice.speak.responseDelay.none": "\u65E0\u5EF6\u8FDF",
+  "dsh-live-voice.speak.responseDelay.oneAndHalf": "1.5 \u79D2",
   "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",
   "dsh-live-voice.speak.segmentGap.label": "\u8BED\u97F3\u7247\u6BB5\u4E4B\u95F4\u7684\u505C\u987F",
   "dsh-live-voice.speak.status.paused": "\u8BED\u97F3\u5DF2\u6682\u505C",
@@ -5050,6 +5212,19 @@ function SpeechStatusBar({ controller }) {
   ));
 }
 
+// src/modules/recognition/transcriptionContext.ts
+var TRANSCRIPTION_CONTEXT_MAX_LENGTH = 1e3;
+function transcriptionContext(value) {
+  if (value === void 0 || value === null) return "";
+  if (typeof value !== "string" || value.length > TRANSCRIPTION_CONTEXT_MAX_LENGTH || value.includes("\0"))
+    throw new Error("Transcription context must contain at most 1000 characters without NUL.");
+  return value.trim();
+}
+function transcriptionContextHeader(value) {
+  const text = transcriptionContext(value);
+  return text ? { "x-dlv-transcription-context": encodeURIComponent(text) } : {};
+}
+
 // src/modules/core/speechDefaults.ts
 var notices = {
   en: "You can see the code in our conversation history.",
@@ -5107,6 +5282,7 @@ var defaultQwenVoice = qwenVoices[0].value;
 var defaultAgentVoiceContext = `Live Voice output is active. Your entire user-facing response will be spoken aloud.
 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.
 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.`;
+var assistantSpeechDelayOptions = Object.freeze([0.6, 1.5, 2, 3, 4, 5]);
 var autoSendDelayOptions = Object.freeze([0.6, 0.8, 1, 2, 3, 4, 5, 6]);
 var recordingBitrateOptions = Object.freeze([24e3, 32e3, 48e3, 64e3, 96e3]);
 var defaultSettings = Object.freeze({
@@ -5142,10 +5318,11 @@ var defaultSettings = Object.freeze({
   sendingMode: "manual",
   dshBusyEnterBehavior: "queue",
   autoSendDelaySeconds: 4,
-  assistantSpeechDelaySeconds: 0,
+  assistantSpeechDelaySeconds: 1.5,
   mode: "speaker",
   lang: "pt-BR",
   recognitionLang: "pt-BR",
+  recognitionContext: "",
   voice: "",
   inputDeviceId: "",
   outputDeviceId: "",
@@ -5207,6 +5384,7 @@ function normalizeSettings(value) {
     browserQwenModel: "qwen-0.6b",
     browserAsrDevice: source.browserAsrDevice === "wasm" ? "wasm" : "webgpu",
     recognitionProcessLocally: typeof source.recognitionProcessLocally === "boolean" ? source.recognitionProcessLocally : defaultSettings.recognitionProcessLocally,
+    recognitionContext: typeof source.recognitionContext === "string" ? source.recognitionContext.replaceAll("\0", "").slice(0, TRANSCRIPTION_CONTEXT_MAX_LENGTH).trim() : defaultSettings.recognitionContext,
     recognitionAutoInstall: typeof source.recognitionAutoInstall === "boolean" ? source.recognitionAutoInstall : defaultSettings.recognitionAutoInstall,
     voiceDetectionEngine: ["energy", "silero"].includes(source.voiceDetectionEngine) ? source.voiceDetectionEngine : defaultSettings.voiceDetectionEngine,
     voiceDetectionPreset: source.voiceDetectionPreset === "custom" || Object.hasOwn(voiceDetectionPresets, source.voiceDetectionPreset) ? source.voiceDetectionPreset : defaultSettings.voiceDetectionPreset,
@@ -5244,7 +5422,9 @@ function normalizeSettings(value) {
     dshBusyEnterBehavior: ["queue", "steer"].includes(source.dshBusyEnterBehavior) ? source.dshBusyEnterBehavior : defaultSettings.dshBusyEnterBehavior,
     sendingMode: source.sendingMode === "automatic" ? "queue" : ["manual", "queue", "steer"].includes(source.sendingMode) ? source.sendingMode : defaultSettings.sendingMode,
     autoSendDelaySeconds: autoSendDelayOptions.includes(source.autoSendDelaySeconds) ? source.autoSendDelaySeconds : defaultSettings.autoSendDelaySeconds,
-    assistantSpeechDelaySeconds: Number.isInteger(source.assistantSpeechDelaySeconds) && source.assistantSpeechDelaySeconds >= 0 && source.assistantSpeechDelaySeconds <= 4 ? source.assistantSpeechDelaySeconds : defaultSettings.assistantSpeechDelaySeconds,
+    assistantSpeechDelaySeconds: assistantSpeechDelayOptions.includes(
+      source.assistantSpeechDelaySeconds
+    ) ? 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,
@@ -6869,8 +7049,23 @@ function cancellable(promise, signal) {
   });
 }
 var VoiceCoordinator = class _VoiceCoordinator {
-  constructor({ recognition, engines, meter, composer, settings = {}, persistSettings }) {
-    Object.assign(this, { recognition, engines, meter, composer, persistSettings });
+  constructor({
+    recognition,
+    engines,
+    meter,
+    composer,
+    settings = {},
+    persistSettings,
+    recordingTimeline
+  }) {
+    Object.assign(this, {
+      recognition,
+      engines,
+      meter,
+      composer,
+      persistSettings,
+      recordingTimeline
+    });
     this.listeners = /* @__PURE__ */ new Set();
     this.transcript = new TranscriptDraft();
     this.epoch = 0;
@@ -6896,6 +7091,7 @@ var VoiceCoordinator = class _VoiceCoordinator {
       recognizing: false,
       pendingTranscriptions: 0,
       pendingTranscriptionRequests: void 0,
+      pendingDeliveryTranscriptions: void 0,
       muted: normalizeSettings(settings).microphoneEnabled === false,
       speaking: false,
       paused: false,
@@ -7240,14 +7436,20 @@ var VoiceCoordinator = class _VoiceCoordinator {
           onResult: (result) => {
             if (valid()) this.onResult(result);
           },
-          onProcessingChange: ({ pending = 0, transcribing } = {}) => {
+          onProcessingChange: ({ pending = 0, transcribing, deliveryPending } = {}) => {
             if (!valid()) return;
             const count = Number.isSafeInteger(pending) && pending >= 0 ? pending : 0;
             const requests = Number.isSafeInteger(transcribing) && transcribing >= 0 ? transcribing : count;
-            this.patch({ pendingTranscriptions: count, pendingTranscriptionRequests: requests });
-            if (count > 0) this.cancelAutoSend({ preserveIntent: true });
-            else if (this.holdToTalkRelease) this._finishHoldToTalk();
-            else this.maybeScheduleAutoSend();
+            const delivery = Number.isSafeInteger(deliveryPending) && deliveryPending >= 0 ? deliveryPending : count;
+            this.patch({
+              pendingTranscriptions: count,
+              pendingTranscriptionRequests: requests,
+              pendingDeliveryTranscriptions: delivery
+            });
+            if (delivery > 0) this.cancelAutoSend({ preserveIntent: true });
+            else if (this.holdToTalkRelease) {
+              if (count === 0) this._finishHoldToTalk();
+            } else this.maybeScheduleAutoSend();
             if (count === 0) this._drain();
           },
           onActivity: (active) => {
@@ -7448,15 +7650,18 @@ var VoiceCoordinator = class _VoiceCoordinator {
     this._handleSpeechInterruption(this._speechActivity());
     this._drain();
   }
+  deliveryPending() {
+    return this.snapshot.pendingDeliveryTranscriptions ?? this.snapshot.pendingTranscriptions;
+  }
   maybeScheduleAutoSend(draft = this.autoSendIntent) {
     if (draft == null || draft !== this.composer.getDraft()) return;
     if (this.autoSendTimer !== null) return;
-    if (this.snapshot.settings.sendingMode === "manual" || this.snapshot.muted || this.composer.canAutoSend?.() === false || this.snapshot.recognizing || this.snapshot.pendingTranscriptions > 0)
+    if (this.snapshot.settings.sendingMode === "manual" || this.snapshot.muted || this.composer.canAutoSend?.() === false || this.snapshot.recognizing || this.deliveryPending() > 0)
       return;
     this.scheduleAutoSend(draft);
   }
   refreshDeliveryReadiness() {
-    if (this.composer.canAutoSend?.() === false || this.snapshot.recognizing || this.snapshot.pendingTranscriptions > 0)
+    if (this.composer.canAutoSend?.() === false || this.snapshot.recognizing || this.deliveryPending() > 0)
       this.cancelAutoSend({ preserveIntent: true });
     else this.maybeScheduleAutoSend();
   }
@@ -7513,6 +7718,7 @@ var VoiceCoordinator = class _VoiceCoordinator {
       recognizing: false,
       pendingTranscriptions: 0,
       pendingTranscriptionRequests: 0,
+      pendingDeliveryTranscriptions: 0,
       starting: false
     });
     return this._input(async () => {
@@ -7673,6 +7879,7 @@ var VoiceCoordinator = class _VoiceCoordinator {
     this.queue = [];
   }
   async stopSpeech(resumeListening = true) {
+    this.activeRecordingTimeline = null;
     this._cancelAssistantSpeechTimer();
     this.assistantSpeechNotBefore = 0;
     this._cancelInterruptionResume();
@@ -7780,8 +7987,10 @@ var VoiceCoordinator = class _VoiceCoordinator {
     try {
       const result = await this.engines[this.snapshot.speechEngine || this.snapshot.settings.engine].pause();
       if (result !== false) this._cancelSpeechPrefetch();
-      if (epoch === this.speechEpoch && control === this.controlEpoch && this.snapshot.speaking && result !== false)
+      if (epoch === this.speechEpoch && control === this.controlEpoch && this.snapshot.speaking && result !== false) {
         this.patch({ paused: true });
+        this.activeRecordingTimeline?.("assistant-pause");
+      }
     } catch (error) {
       if (epoch === this.speechEpoch) this.patch({ error: message(error) });
     }
@@ -7821,6 +8030,7 @@ var VoiceCoordinator = class _VoiceCoordinator {
       const result = await this.engines[this.snapshot.speechEngine || this.snapshot.settings.engine].resume();
       if (epoch === this.speechEpoch && control === this.controlEpoch && result !== false) {
         this.patch({ paused: false });
+        this.activeRecordingTimeline?.("assistant-resume");
         this._prefetchSpeech();
       }
     } catch (error) {
@@ -7864,6 +8074,10 @@ var VoiceCoordinator = class _VoiceCoordinator {
     const controller = new AbortController();
     this.speechController = controller;
     this.speechLaunchPending = true;
+    const timeline = this.activeRecordingTimeline ?? this.recordingTimeline?.();
+    this.activeRecordingTimeline = timeline;
+    let playbackStarted = false;
+    let playbackFinished = false;
     const engine = this.engines[this.snapshot.settings.engine];
     this._resetStaleSpeechItem(item, this._speechPreparationKey());
     item.engine = engine;
@@ -7888,6 +8102,7 @@ var VoiceCoordinator = class _VoiceCoordinator {
       if (this.speechStopError) throw this.speechStopError;
       if (epoch !== this.speechEpoch || this.disposed) return;
       if (this.snapshot.settings.mode === "speaker") {
+        if (this.snapshot.listening) timeline?.("capture-gated");
         await this.stopListening();
         if (this.inputReleaseError) throw this.inputReleaseError;
       }
@@ -7898,7 +8113,11 @@ var VoiceCoordinator = class _VoiceCoordinator {
       if (queued?.prepareError) throw queued.prepareError;
       const options = {
         onPlaybackStart: () => {
-          if (epoch === this.speechEpoch && !this.disposed) this.patch({ speechLoading: false });
+          if (epoch === this.speechEpoch && !this.disposed && !controller.signal.aborted && !playbackFinished) {
+            if (!playbackStarted) timeline?.("assistant-start");
+            playbackStarted = true;
+            this.patch({ speechLoading: false });
+          }
         },
         voice: this.snapshot.settings.voice || void 0,
         rate: this.snapshot.settings.rate,
@@ -7931,6 +8150,10 @@ var VoiceCoordinator = class _VoiceCoordinator {
         if (error.name !== "AbortError") this.patch({ error: message(error) });
       }
     } finally {
+      playbackFinished = true;
+      if (playbackStarted) timeline?.("assistant-end");
+      if (this.activeRecordingTimeline === timeline && epoch === this.speechEpoch && !this.queue.length)
+        this.activeRecordingTimeline = null;
       if (epoch === this.speechEpoch && !this.disposed) {
         this.patch({
           speaking: false,
@@ -8089,6 +8312,8 @@ var MeetingController = class {
       if (Object.hasOwn(next, "pending")) policy.pendingTranscriptions = next.pending;
       if (Object.hasOwn(next, "transcribing"))
         policy.pendingTranscriptionRequests = next.transcribing;
+      if (Object.hasOwn(next, "deliveryPending"))
+        policy.pendingDeliveryTranscriptions = next.deliveryPending;
       input?.patch(policy);
     }
     this.snapshot = {
@@ -8261,11 +8486,12 @@ var MeetingController = class {
           this.onActivity?.(source, active);
           input.refreshDeliveryReadiness();
         },
-        onProcessingChange: ({ pending, transcribing }) => {
+        onProcessingChange: ({ pending, transcribing, deliveryPending }) => {
           if (!valid()) return;
           const requests = Number.isSafeInteger(transcribing) && transcribing >= 0 ? transcribing : pending;
-          this.patch(source, { pending, transcribing: requests });
-          if (pending) input.cancelAutoSend({ preserveIntent: true });
+          const delivery = Number.isSafeInteger(deliveryPending) && deliveryPending >= 0 ? deliveryPending : pending;
+          this.patch(source, { pending, transcribing: requests, deliveryPending: delivery });
+          if (delivery) input.cancelAutoSend({ preserveIntent: true });
           else if (!input.getSnapshot().muted) input.maybeScheduleAutoSend();
         },
         onError: (error) => {
@@ -8308,12 +8534,19 @@ var MeetingController = class {
       listening: false,
       recognizing: false,
       pendingTranscriptions: 0,
-      pendingTranscriptionRequests: 0
+      pendingTranscriptionRequests: 0,
+      pendingDeliveryTranscriptions: 0
     });
     this.onActivity?.(source, false);
     job?.abort.abort();
     this.transcript.setActive(source, false);
-    this.patch(source, { starting: false, listening: false, pending: 0, transcribing: 0 });
+    this.patch(source, {
+      starting: false,
+      listening: false,
+      pending: 0,
+      transcribing: 0,
+      deliveryPending: 0
+    });
     await job?.engine?.stop();
     await job?.meter?.stop();
   }
@@ -9494,6 +9727,12 @@ var PinDetector = class {
       }
     } else this.candidate = 0;
   }
+  /** Advance past unclassified saved samples without inventing a speech/silence transition. */
+  skipTo(position) {
+    this.at = Math.max(this.at, position);
+    this.candidate = 0;
+    this.quiet = 0;
+  }
   finish() {
     if (this.speaking) {
       this.speaking = false;
@@ -9505,6 +9744,7 @@ var PinDetector = class {
 
 // src/modules/recording/RecordingRecognition.ts
 var asError = (value) => value instanceof Error ? value : new Error(String(value));
+var TRANSCRIPTION_BOUNDARY_MARGIN_MS = 800;
 var uid = () => crypto.randomUUID();
 async function startRecordingRecognition(engine, options, capture) {
   const { signal, lang = engine.lang, onResult, onActivity, onError, onProcessingChange } = options;
@@ -9513,8 +9753,8 @@ async function startRecordingRecognition(engine, options, capture) {
     enabled: engine.recordingEnabled !== false,
     keepPrevious: engine.recordingKeepPrevious ?? 1,
     bitrate: engine.recordingBitrate ?? 48e3,
-    preRollMs: engine.recordingPreRollMs ?? 320,
-    postRollMs: engine.recordingPostRollMs ?? 320
+    preRollMs: Math.max(TRANSCRIPTION_BOUNDARY_MARGIN_MS, engine.recordingPreRollMs ?? 320),
+    postRollMs: Math.max(TRANSCRIPTION_BOUNDARY_MARGIN_MS, engine.recordingPostRollMs ?? 320)
   };
   const generation = engine.startGeneration;
   const writer = await engine.recordings.begin({
@@ -9562,6 +9802,7 @@ async function startRecordingRecognition(engine, options, capture) {
     onProcessingChange,
     tail: Promise.resolve(),
     pendingWrites: 0,
+    pendingSpeechWrites: 0,
     lang,
     operation: uid(),
     capturedSamples: 0,
@@ -9574,18 +9815,22 @@ async function startRecordingRecognition(engine, options, capture) {
   };
   engine.session = session;
   const valid = () => engine.session === session && !signal?.aborted;
+  const awaitingTail = [];
   let stopping = false;
+  let recoveryPending = false;
   let captureFailure;
   const captureValid = () => engine.session === session || stopping;
   const notify2 = () => {
     if (valid())
       session.onProcessingChange?.({
-        queued: session.transcriptionQueue.length,
+        queued: awaitingTail.length + session.transcriptionQueue.length,
         active: !!session.activeRequest,
         // Only queued/running speech attempts belong in the visible transcription status.
-        transcribing: session.transcriptionQueue.length + (session.activeRequest ? 1 : 0),
-        // Audio writes and final drain still block delivery/release, even in silence.
-        pending: session.transcriptionQueue.length + (session.activeRequest ? 1 : 0) + session.pendingWrites + (session.finishing ? 1 : 0)
+        transcribing: awaitingTail.length + session.transcriptionQueue.length + (session.activeRequest ? 1 : 0),
+        // Speech-boundary bookkeeping must settle before delivery, but silent PCM writes must not reset it.
+        deliveryPending: awaitingTail.length + session.transcriptionQueue.length + (session.activeRequest ? 1 : 0) + session.pendingSpeechWrites + (session.finishing || recoveryPending ? 1 : 0),
+        // All writes still protect release, playback admission and page leave.
+        pending: awaitingTail.length + session.transcriptionQueue.length + (session.activeRequest ? 1 : 0) + session.pendingWrites + (session.finishing || recoveryPending ? 1 : 0)
       });
   };
   const failure = (error) => {
@@ -9594,14 +9839,16 @@ async function startRecordingRecognition(engine, options, capture) {
     onError?.(new Error(engine.recordingErrorMessage(error)));
     void engine.stop();
   };
-  const serialize = (task, preserve = false) => {
+  const serialize = (task, preserve = false, speech = false) => {
     session.pendingWrites++;
+    if (speech) session.pendingSpeechWrites++;
     notify2();
     const work = session.tail.then(async () => {
       if (preserve || valid()) await task();
     });
     session.tail = work.catch(failure).finally(() => {
       session.pendingWrites--;
+      if (speech) session.pendingSpeechWrites--;
       notify2();
     });
     return work;
@@ -9621,19 +9868,20 @@ async function startRecordingRecognition(engine, options, capture) {
     if (through - Math.max(liveSpeechStart, lastSplitQueued) < max) return;
     const end = Math.max(liveSpeechStart, lastSplitQueued) + max;
     lastSplitQueued = end;
-    void serialize(async () => {
-      const speech = [...writer.session.pins].reverse().find((p) => p.kind === "speech-start");
-      if (speech && valid()) await queueSpeech(speech, end);
-    }).catch(() => {
+    void serialize(
+      async () => {
+        const speech = [...writer.session.pins].reverse().find((p) => p.kind === "speech-start");
+        if (speech && valid()) await queueSpeech(speech, end);
+      },
+      false,
+      true
+    ).catch(() => {
     });
   };
   const queueSpeech = async (pin2, end) => {
     for (let at2 = Math.max(pin2.atSample, queuedThrough); at2 < end; at2 += max) {
       const requestedStart = Math.max(0, at2 - policy.preRollMs * rate / 1e3);
-      const requestedEnd = Math.min(
-        writer.session.capturedSamples,
-        Math.min(end, at2 + max) + policy.postRollMs * rate / 1e3
-      );
+      const requestedEnd = Math.min(end, at2 + max) + policy.postRollMs * rate / 1e3;
       const attempt = {
         id: uid(),
         segmentId: pin2.id,
@@ -9645,13 +9893,26 @@ async function startRecordingRecognition(engine, options, capture) {
         createdAt: (/* @__PURE__ */ new Date()).toISOString(),
         status: "pending"
       };
+      awaitingTail.push(attempt);
+      notify2();
+    }
+    queuedThrough = Math.max(queuedThrough, end);
+    await releaseCapturedRanges();
+  };
+  async function releaseCapturedRanges() {
+    if (!valid()) return;
+    const captured = writer.session.capturedSamples;
+    while (awaitingTail.length && (session.finished || awaitingTail[0].endSample <= captured)) {
+      const pending = awaitingTail[0];
+      const attempt = { ...pending, endSample: Math.min(pending.endSample, captured) };
       await writer.attempt(attempt);
+      if (!valid()) return;
+      awaitingTail.shift();
       session.transcriptionQueue.push(attempt);
       notify2();
     }
-    queuedThrough = Math.max(queuedThrough, end);
     void drain();
-  };
+  }
   const pin = (kind, at2, confirmed) => {
     const startSample = Math.min(
       session.capturedSamples,
@@ -9665,14 +9926,18 @@ async function startRecordingRecognition(engine, options, capture) {
       liveSpeechStart = startSample;
       lastSplitQueued = startSample;
     } else liveSpeechStart = null;
-    void serialize(async () => {
-      const previous = writer.session.pins.at(-1);
-      if (kind === "speech-start" && previous?.kind === "silence-start" && previous.atSample > 0 && startSample - previous.atSample >= paragraphSamples)
-        paragraphStarts.add(startSample);
-      await writer.pin(kind, startSample, confirmedSample);
-      if (valid() && kind === "silence-start" && previous?.kind === "speech-start")
-        await queueSpeech(previous, startSample);
-    }, true).catch(() => {
+    void serialize(
+      async () => {
+        const previous = writer.session.pins.at(-1);
+        if (kind === "speech-start" && previous?.kind === "silence-start" && previous.atSample > 0 && startSample - previous.atSample >= paragraphSamples)
+          paragraphStarts.add(startSample);
+        await writer.pin(kind, startSample, confirmedSample);
+        if (valid() && kind === "silence-start" && previous?.kind === "speech-start")
+          await queueSpeech(previous, startSample);
+      },
+      true,
+      true
+    ).catch(() => {
     });
   };
   const detector = new PinDetector({
@@ -9714,7 +9979,8 @@ async function startRecordingRecognition(engine, options, capture) {
                 "content-type": "audio/wav",
                 "x-dlv-client-id": session.operation,
                 "x-dlv-operation-id": attempt.id,
-                "x-dlv-language": lang
+                "x-dlv-language": lang,
+                ...transcriptionContextHeader(engine.recognitionContext)
               },
               credentials: "same-origin",
               body: capture.encode(pcm, rate),
@@ -9768,7 +10034,114 @@ async function startRecordingRecognition(engine, options, capture) {
       return;
     }
     const copy = pcm.slice();
-    void serialize(() => writer.append(copy), true).catch(() => {
+    void serialize(async () => {
+      await writer.append(copy);
+      await releaseCapturedRanges();
+    }, true).catch(() => {
+    });
+  };
+  let vadGeneration = 0;
+  let restartAttempts = 0;
+  let recovering = null;
+  let deferredFailure = null;
+  const sileroValid = () => valid() && !stopping && !session.finished && !session.finishing;
+  const releaseSilero = () => {
+    ++vadGeneration;
+    const previousCapture = session.capture;
+    const previousVad = session.vad;
+    session.capture = null;
+    session.vad = null;
+    try {
+      previousCapture?.release();
+    } finally {
+      previousVad?.release();
+    }
+  };
+  const terminalVadFailure = (error) => {
+    if (!valid() || stopping || session.finished) return;
+    captureFailure = asError(error).message;
+    onError?.(new Error(engine.vadErrorMessage(error)));
+    void engine.stop();
+  };
+  const installSilero = async () => {
+    const token = ++vadGeneration;
+    let installing = true;
+    let installationFailure;
+    const failed = (error) => {
+      if (token !== vadGeneration || !sileroValid()) return;
+      if (installing) installationFailure = error;
+      else recoverSilero(error);
+    };
+    const vad = engine.vadFactory({
+      onProbability: ({ probability, pcm }) => {
+        if (token !== vadGeneration || !captureValid() || session.finished) return;
+        restartAttempts = 0;
+        detector.frame(probability, pcm.length);
+        checkLiveSplit();
+      },
+      onError: failed,
+      onReset: () => failed(new Error("VAD analysis interrupted."))
+    });
+    session.vad = vad;
+    try {
+      const attached = await engine.vadCapture(
+        engine.meter.context,
+        engine.meter.source,
+        vad,
+        (pcm, sampleRate) => {
+          if (token === vadGeneration) ingest(pcm, sampleRate);
+        }
+      );
+      if (token !== vadGeneration || !sileroValid()) {
+        attached.release();
+        vad.release();
+        throw Object.assign(new Error("Cancelled"), { name: "AbortError" });
+      }
+      session.capture = attached;
+      if (installationFailure) throw installationFailure;
+    } finally {
+      installing = false;
+    }
+  };
+  const retrySilero = async (initialError) => {
+    let error = initialError;
+    while (sileroValid()) {
+      if (error) {
+        releaseSilero();
+        if (restartAttempts >= 2) throw error;
+        restartAttempts++;
+        detector.skipTo(Math.round(session.capturedSamples * 16e3 / rate));
+        await Promise.resolve();
+      }
+      if (!sileroValid()) return;
+      try {
+        await installSilero();
+        return;
+      } catch (failure2) {
+        if (!sileroValid() || asError(failure2).name === "AbortError") return;
+        error = failure2;
+      }
+    }
+  };
+  const recoverSilero = (error) => {
+    if (!sileroValid()) return;
+    if (recovering) {
+      ++vadGeneration;
+      deferredFailure = error;
+      return;
+    }
+    ++vadGeneration;
+    recoveryPending = true;
+    notify2();
+    recovering = Promise.resolve().then(() => retrySilero(error)).catch(terminalVadFailure).finally(() => {
+      recovering = null;
+      const pending = deferredFailure;
+      deferredFailure = null;
+      if (pending) recoverSilero(pending);
+      else {
+        recoveryPending = false;
+        notify2();
+      }
     });
   };
   session.finishRecording = () => {
@@ -9776,10 +10149,12 @@ async function startRecordingRecognition(engine, options, capture) {
     session.finishing = true;
     session.finishPromise = Promise.resolve().then(async () => {
       try {
+        await recovering;
         await session.capture?.finish();
         if (!valid()) return;
         session.finished = true;
         detector.finish();
+        await serialize(releaseCapturedRanges);
         await session.tail;
         await writer.finish();
         session.vad?.release();
@@ -9797,6 +10172,8 @@ async function startRecordingRecognition(engine, options, capture) {
   };
   session.stopRecording = async () => {
     stopping = true;
+    awaitingTail.length = 0;
+    await recovering;
     session.activeRequest?.abort();
     if (session.finishing) session.capture?.release();
     else {
@@ -9826,31 +10203,7 @@ async function startRecordingRecognition(engine, options, capture) {
   signal?.addEventListener("abort", session.abort, { once: true });
   try {
     if (capture.silero) {
-      session.vad = engine.vadFactory({
-        onProbability: ({ probability, pcm }) => {
-          if (captureValid() && !session.finished) {
-            detector.frame(probability, pcm.length);
-            checkLiveSplit();
-          }
-        },
-        onError: (error) => {
-          captureFailure = error.message;
-          if (valid()) onError?.(new Error(engine.vadErrorMessage(error)));
-          void engine.stop();
-        },
-        onReset: () => {
-          captureFailure = "VAD analysis interrupted.";
-          if (valid())
-            onError?.(new Error(engine.vadErrorMessage(new Error("VAD analysis interrupted."))));
-          void engine.stop();
-        }
-      });
-      session.capture = await engine.vadCapture(
-        engine.meter.context,
-        engine.meter.source,
-        session.vad,
-        ingest
-      );
+      await retrySilero();
     } else {
       const processor = engine.meter.context.createScriptProcessor(4096, 1, 1);
       const gain = engine.meter.context.createGain();
@@ -9936,6 +10289,7 @@ var WhisperHttpRecognitionEngine = class {
     voiceDetectionPreset = defaultSettings.voiceDetectionPreset,
     voiceDetectionCustomSilenceMs = defaultSettings.voiceDetectionCustomSilenceMs,
     voiceDetectionCustomParagraphMs = defaultSettings.voiceDetectionCustomParagraphMs,
+    recognitionContext = "",
     maxUtteranceSeconds = 60,
     recordings,
     inputSource = "microphone",
@@ -9962,6 +10316,7 @@ var WhisperHttpRecognitionEngine = class {
     this.voiceDetectionPreset = voiceDetectionPreset;
     this.voiceDetectionCustomSilenceMs = voiceDetectionCustomSilenceMs;
     this.voiceDetectionCustomParagraphMs = voiceDetectionCustomParagraphMs;
+    this.recognitionContext = recognitionContext;
     this.maxUtteranceSeconds = maxUtteranceSeconds;
     Object.assign(this, {
       recordings: recordings ?? createRecordingRepository(),
@@ -10053,7 +10408,13 @@ var WhisperHttpRecognitionEngine = class {
     this.session = null;
     session.signal?.removeEventListener("abort", session.abort);
     await session.stopRecording();
-    session.onProcessingChange?.({ queued: 0, active: false, transcribing: 0, pending: 0 });
+    session.onProcessingChange?.({
+      queued: 0,
+      active: false,
+      transcribing: 0,
+      deliveryPending: 0,
+      pending: 0
+    });
   }
 };
 
@@ -10532,8 +10893,8 @@ function createDeviceOptions(devices, kind, systemDefault, fallback) {
 var import_react36 = __toESM(require("react"), 1);
 
 // src/modules/settings/services/releases.ts
-var CURRENT_VERSION = "0.6.0";
-var TESTED_DSH_VERSION = "0.2.1-alpha.2";
+var CURRENT_VERSION = "0.6.2";
+var TESTED_DSH_VERSION = "0.2.1-rc.2";
 var REPOSITORY_URL = "https://github.com/victorwads/dsh-live-voice";
 var RELEASES_URL = `${REPOSITORY_URL}/releases`;
 var LATEST_RELEASE_API_URL = "https://api.github.com/repos/victorwads/dsh-live-voice/releases/latest";
@@ -10761,7 +11122,6 @@ function VoiceModeSettings({
 
 // src/modules/settings/sections/conversation/ConversationDelaySettings.tsx
 var import_react40 = __toESM(require("react"), 1);
-var DELAYS = [0, 1, 2, 3, 4];
 function ConversationDelaySettings({ settings, updateSettings }) {
   const { scoped: commons } = useLanguage((ctx) => ctx.commons);
   const { scoped: speak } = useLanguage((ctx) => ctx.speak);
@@ -10770,10 +11130,12 @@ function ConversationDelaySettings({ settings, updateSettings }) {
     {
       label: speak.responseDelay.label(),
       hint: speak.responseDelay.help(),
-      value: String(settings.assistantSpeechDelaySeconds ?? 0),
-      options: DELAYS.map((seconds) => ({
+      value: String(
+        settings.assistantSpeechDelaySeconds ?? defaultSettings.assistantSpeechDelaySeconds
+      ),
+      options: assistantSpeechDelayOptions.map((seconds) => ({
         value: String(seconds),
-        label: seconds === 0 ? speak.responseDelay.none() : seconds === 1 ? commons.second() : commons.seconds({ seconds })
+        label: seconds < 1 ? commons.milliseconds({ milliseconds: seconds * 1e3 }) : seconds === 1.5 ? speak.responseDelay.oneAndHalf() : commons.seconds({ seconds })
       })),
       onChange: (event) => updateSettings({ assistantSpeechDelaySeconds: Number(event.target.value) })
     }
@@ -11518,7 +11880,18 @@ function RecognitionEngineSettings({
       settings,
       updateSettings
     }
-  ), 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 })));
+  ), ["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(
+    TextAreaField,
+    {
+      label: recognition.context.label(),
+      hint: recognition.context.hint(),
+      placeholder: recognition.context.placeholder(),
+      value: settings.recognitionContext || "",
+      rows: 3,
+      maxLength: TRANSCRIPTION_CONTEXT_MAX_LENGTH,
+      onCommit: (recognitionContext) => updateSettings({ recognitionContext })
+    }
+  )));
 }
 
 // src/modules/settings/sections/recognition/RecognitionFilterSettings.tsx
@@ -13057,7 +13430,9 @@ function createPluginUpdate({
         if (!result?.ok) return patch({ ...state, phase: "uncertain", problem: "uncertain" });
         const value = result.value;
         if (value.application === "failed" || value.application === "cancelled")
-          return fail(value.pendingBuilds?.length ? "scripts" : "failed");
+          return fail(
+            value.error?.code === "incompatible-version" || value.packageResult?.incompatible?.length ? "incompatible" : value.pendingBuilds?.length || value.packageResult?.kind === "build-blocked" ? "scripts" : ["not-found", "no-matching-version"].includes(value.packageResult?.kind) ? "delayed" : "failed"
+          );
         if (value.bundle !== "dsh-live-voice" || value.version !== state.version || !["applied", "restart-required"].includes(value.application))
           return patch({ ...state, phase: "uncertain", problem: "uncertain" });
         patch({ ...state, phase: value.application === "restart-required" ? "restart" : "done" });
@@ -13270,7 +13645,7 @@ async function removeBrowserModel(key, storage = caches) {
 }
 
 // src/modules/recognition/local/assets.ts
-var browserAsrAssetBase = () => "/api/dsh-live-voice/asr/" + encodeURIComponent(true ? "0.6.0" : "test") + "/";
+var browserAsrAssetBase = () => "/api/dsh-live-voice/asr/" + encodeURIComponent(true ? "0.6.2" : "test") + "/";
 
 // src/modules/recognition/local/runtimeIdentity.ts
 var workerProtocol = "asr-worker-v3-ort-1.31.0-dev.20260914-8d85527a0";
@@ -13738,7 +14113,7 @@ var BrowserLocalRecognitionEngine = class extends WhisperHttpRecognitionEngine {
   message;
   release;
   constructor({ manager, modelKey, device = "webgpu", message: message2, ...options }) {
-    const margins = ((options.recordingPreRollMs ?? 320) + (options.recordingPostRollMs ?? 320)) / 1e3;
+    const margins = (Math.max(TRANSCRIPTION_BOUNDARY_MARGIN_MS, options.recordingPreRollMs ?? 320) + Math.max(TRANSCRIPTION_BOUNDARY_MARGIN_MS, options.recordingPostRollMs ?? 320)) / 1e3;
     super({
       ...options,
       maxUtteranceSeconds: Math.max(
@@ -13980,6 +14355,15 @@ function apply(ctx) {
   const settingsControllers = /* @__PURE__ */ new Set();
   let disposed = false;
   const recordings = createRecordingRepository();
+  const recordingTimeline = () => {
+    const append = recordings.timeline();
+    return (kind) => {
+      void append(kind).catch(() => {
+        for (const entry of controllers.values())
+          entry.controller.patch({ error: t("dsh-live-voice.recording.error.failed") });
+      });
+    };
+  };
   const localModels = new LocalModelManager();
   void localModels.refresh();
   const leaveGuard = createTranscriptionLeaveGuard(window);
@@ -14008,7 +14392,7 @@ function apply(ctx) {
       },
       updatePlaybackSettings: (next) => meetingEntry?.controller.updateSettings(next),
       handleQuestionResult: (result) => meetingEntry?.controller.composer.handleQuestionResult?.(result),
-      canAutoSend: () => !!meetingEntry && !meetingEntry.controller.getSnapshot().recognizing && !meetingEntry.controller.getSnapshot().pendingTranscriptions
+      canAutoSend: () => !!meetingEntry && !meetingEntry.controller.getSnapshot().recognizing && !meetingEntry.controller.deliveryPending()
     },
     settings: () => preferences.getSnapshot(),
     savePreferences: (controller, next) => savePreferences(controller, next),
@@ -14142,6 +14526,7 @@ function apply(ctx) {
     "browserQwenModel",
     "browserAsrDevice",
     "recognitionLang",
+    "recognitionContext",
     "inputDeviceId",
     "recognitionProcessLocally",
     "recognitionAutoInstall",
@@ -14161,6 +14546,7 @@ function apply(ctx) {
   const changesRecognition = (next) => Object.keys(next).some((key) => recognitionSettingKeys.has(key));
   const recordingOptions = (settings, inputSource) => ({
     recordings,
+    recognitionContext: settings.recognitionContext,
     inputSource,
     sourceId: inputSource,
     sourceLabel: inputSource === "shared" ? settings.sharedSourceName : settings.micSourceName,
@@ -14361,13 +14747,14 @@ function apply(ctx) {
     const meter = new MicrophoneMeter();
     const recognition = recognitionFor(settings, meter);
     entry.controller = new VoiceCoordinator({
+      recordingTimeline,
       persistSettings: (next) => savePreferences(entry.controller, next),
       recognition,
       engines: { browser: engineBrowser, say: engineSay, "qwen-http": engineQwen },
       meter,
       composer: {
         getDraft: () => entry.draft,
-        canAutoSend: () => !meeting.getSnapshot().shared.pending && !meeting.input().getSnapshot().recognizing,
+        canAutoSend: () => !meeting.input().deliveryPending() && !meeting.input().getSnapshot().recognizing,
         appendFinal: (text, startedAt, paragraphBefore) => {
           meeting.cancelDelivery();
           if (meetingEntry === entry && (entry.meeting?.getSnapshot().shared.listening || entry.controller.getSnapshot().timestamps)) {
@@ -14764,6 +15151,7 @@ function apply(ctx) {
       const meter = new MicrophoneMeter();
       const recognition = recognitionFor(settings, meter);
       const c = new VoiceCoordinator({
+        recordingTimeline,
         persistSettings: (next) => savePreferences(c, next),
         recognition,
         engines: {