Updates

daemon/voice-loop.js

@@ -198,6 +198,7 @@ export class VoiceLoop extends EventEmitter {
   async speak(text) {
     const generation = this._generation || 0;
     if (this.tts && !this.status.tts && typeof this.tts.start === 'function') await this.ensureTts();
+    if (generation !== this._generation) return;
     if (!isSpeakable(text) || !this.tts?.speak || !this.status.tts) {
       this._releaseSpeaking();
       return;
@@ -209,32 +210,41 @@ export class VoiceLoop extends EventEmitter {
       return;
     }
     this._speechQueue = this._speechQueue.catch(() => {}).then(async () => {
+      let pending;
+      const prepare = (sentence) => Promise.resolve().then(async () => {
+        if (generation !== this._generation) return {};
+        const started = this.now();
+        const audio = await this.tts.speak(speakableForTts(sentence));
+        this.metrics.synthesis(this.now() - started, audio.samples.length * 1000 / audio.sampleRate);
+        return { audio };
+      }).catch((error) => ({ error }));
       try {
         if (generation !== this._generation) return;
-        for (const sentence of sentences) {
+        this.isSpeaking = true;
+        this.wake.pause();
+        pending = prepare(sentences[0]);
+        for (let i = 0; i < sentences.length; i++) {
+          const { audio, error } = await pending;
           if (generation !== this._generation) return;
-          await this.speakSentence(sentence, generation);
+          if (error) throw error;
+          // Only one sentence ahead: overlap synthesis with playback, without
+          // parallel TTS requests or buffering an entire reply's audio.
+          pending = i + 1 < sentences.length ? prepare(sentences[i + 1]) : null;
+          await this.playback.play(audio.samples, audio.sampleRate);
+          try { this.daemon?.emit('SpeakingLevel', 0); } catch {}
         }
       } finally {
-        if (generation === this._generation) this._releaseSpeaking();
+        // A failed/interrupted playback may leave one synthesis in flight.
+        // Drain it before model teardown or the next queued utterance.
+        await pending;
+        if (generation === this._generation) {
+          this.cooldownUntil = this.now() + this.cooldownMs;
+          this._releaseSpeaking();
+        }
       }
     });
     return this._speechQueue;
   }
-  async speakSentence(text, generation) {
-    this.isSpeaking = true;
-    this.wake.pause();
-    try {
-      const audio = await this.tts.speak(speakableForTts(text));
-      if (generation !== this._generation) return;
-      await this.playback.play(audio.samples, audio.sampleRate);
-      try { this.daemon?.emit('SpeakingLevel', 0); } catch {}
-    } finally {
-      this.isSpeaking = false;
-      this.cooldownUntil = this.now() + this.cooldownMs;
-      this.wake.resume();
-    }
-  }
 }
 
 export { SentenceBuffer };