feat: 실시간 음성대화 GPU 엔진(whisper+XTTS) + Ollama 쿼터 자동 폴백

- voice_engine.py: faster-whisper(STT)+XTTS-v2(TTS)를 상시 로드해 로컬에서 서빙하는 WebSocket 엔진 (배치/스트리밍 프로토콜) - routes-voice-realtime.ts: 브라우저 WS를 voice_engine.py로 그대로 프록시 - voice-call.js/worklets: 실시간 연속 대화 모드(VAD 기반 발화 감지, barge-in), 마크다운/표/URL을 정리하고 읽는 sanitizeForSpeech 포함 - tts.ts/stt.ts: xtts_gpu/whisper_gpu provider 분기 추가 - ollama-client.ts/factory.ts: Ollama Cloud 세션 쿼터 초과 시 설정된 fallback 모델로 자동 재시도 Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
2026-07-09 18:05:55 +09:00
parent 94caf1c4da
commit 22c692fec3
14 changed files with 1153 additions and 14 deletions
@@ -0,0 +1,81 @@
+// AudioWorkletProcessors for the real-time voice call mode.
+// Kept deliberately dumb: capture does downsampling + chunking + RMS only,
+// playback does ring-buffer PCM output only. All turn-taking / VAD / WS
+// logic lives in voice-call.js on the main thread.
+
+class MicCaptureProcessor extends AudioWorkletProcessor {
+  constructor() {
+    super();
+    this.targetRate = 16000;
+    this.srcRate = sampleRate; // AudioWorkletGlobalScope — native context rate
+    this.ratio = this.srcRate / this.targetRate;
+    this.acc = 0;            // fractional resample accumulator
+    this.chunk = [];         // pending int16 samples for the current outgoing chunk
+    this.chunkSamples = Math.round(this.targetRate * 0.1); // ~100ms per chunk
+  }
+
+  process(inputs) {
+    const input = inputs[0];
+    if (!input || !input[0]) return true;
+    const ch = input[0];
+
+    for (let i = 0; i < ch.length; i++) {
+      this.acc += 1;
+      if (this.acc >= this.ratio) {
+        this.acc -= this.ratio;
+        const s = Math.max(-1, Math.min(1, ch[i]));
+        this.chunk.push(s < 0 ? s * 0x8000 : s * 0x7fff);
+        if (this.chunk.length >= this.chunkSamples) {
+          const int16 = new Int16Array(this.chunk);
+          let sumSq = 0;
+          for (let j = 0; j < int16.length; j++) { const v = int16[j] / 32768; sumSq += v * v; }
+          const rms = Math.sqrt(sumSq / int16.length);
+          this.port.postMessage({ type: 'audio', pcm: int16, rms }, [int16.buffer]);
+          this.chunk = [];
+        }
+      }
+    }
+    return true;
+  }
+}
+registerProcessor('mic-capture-processor', MicCaptureProcessor);
+
+class PcmPlayerProcessor extends AudioWorkletProcessor {
+  constructor() {
+    super();
+    this.queue = [];   // array of Float32Array
+    this.qi = 0;        // read offset within queue[0]
+    this.port.onmessage = (e) => {
+      const msg = e.data;
+      if (msg.type === 'push') {
+        const int16 = msg.pcm;
+        const f32 = new Float32Array(int16.length);
+        for (let i = 0; i < int16.length; i++) f32[i] = int16[i] / 32768;
+        this.queue.push(f32);
+      } else if (msg.type === 'clear') {
+        this.queue = [];
+        this.qi = 0;
+      }
+    };
+  }
+
+  process(_inputs, outputs) {
+    const out = outputs[0][0];
+    if (!out) return true;
+    let oi = 0;
+    while (oi < out.length) {
+      if (this.queue.length === 0) {
+        out[oi++] = 0;
+        continue;
+      }
+      const cur = this.queue[0];
+      out[oi++] = cur[this.qi++];
+      if (this.qi >= cur.length) {
+        this.queue.shift();
+        this.qi = 0;
+      }
+    }
+    return true;
+  }
+}
+registerProcessor('pcm-player-processor', PcmPlayerProcessor);