agentgui 1.0.179 → 1.0.180

package/lib/speech.js

@@ -8,7 +8,6 @@ import * as pocket from './pocket-sidecar.js';
 const require = createRequire(import.meta.url);
 const __dirname = path.dirname(fileURLToPath(import.meta.url));
 const ROOT = path.dirname(__dirname);
-const DATA_DIR = path.join(ROOT, 'data');
 const AUDIO_EXTENSIONS = ['.wav', '.mp3', '.ogg', '.flac', '.m4a'];
 
 function getVoiceDirs() {
@@ -102,28 +101,11 @@ function getVoices() {
   return [...BASE_VOICES, ...loadCustomVoices()];
 }
 
-const SPEAKER_OFFSETS = { awb: 0, bdl: 1200, clb: 2300, jmk: 3500, ksp: 4700, rms: 5900, slt: 7100 };
-const SPEAKER_EMBEDDINGS_URL = 'https://huggingface.co/datasets/Xenova/speaker_embeddings/resolve/main/spkrec-xvectors-voxceleb.hf';
-const SPEAKER_EMBEDDINGS_PATH = path.join(DATA_DIR, 'speaker_embeddings.bin');
-const DATASET_API = 'https://datasets-server.huggingface.co/rows?dataset=Xenova%2Fspeaker_embeddings&config=default&split=train';
-const SAMPLES_TO_AVERAGE = 30;
-const DEFAULT_EMBEDDING_B64 = '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';
-
 let transformersModule = null;
 let sttPipeline = null;
-let ttsPipeline = null;
-let speakerEmbeddings = null;
-let speakerEmbeddingPipeline = null;
 let sttLoading = false;
-let ttsLoading = false;
-let speakerEmbeddingLoading = false;
-let ttsLoadError = null;
-let ttsLoadErrorTime = 0;
 let sttLoadError = null;
-const voiceEmbeddingsCache = new Map();
 const SAMPLE_RATE_STT = 16000;
-const SAMPLE_RATE_TTS = 16000;
-const TTS_ERROR_RETRY_MS = 30000;
 
 const TTS_CACHE_MAX_BYTES = 10 * 1024 * 1024;
 let ttsCacheBytes = 0;
@@ -145,109 +127,6 @@ function whisperModelPath() {
   return 'onnx-community/whisper-base';
 }
 
-function defaultEmbedding() {
-  const buf = Buffer.from(DEFAULT_EMBEDDING_B64, 'base64');
-  return new Float32Array(new Uint8Array(buf).buffer);
-}
-
-async function ensureSpeakerEmbeddings() {
-  if (speakerEmbeddings) return speakerEmbeddings;
-  if (!fs.existsSync(DATA_DIR)) fs.mkdirSync(DATA_DIR, { recursive: true });
-  if (fs.existsSync(SPEAKER_EMBEDDINGS_PATH)) {
-    const buf = fs.readFileSync(SPEAKER_EMBEDDINGS_PATH);
-    if (buf.length === 2048) {
-      speakerEmbeddings = new Float32Array(new Uint8Array(buf).buffer);
-      return speakerEmbeddings;
-    }
-  }
-  try {
-    const resp = await fetch(SPEAKER_EMBEDDINGS_URL);
-    if (resp.ok) {
-      const data = Buffer.from(await resp.arrayBuffer());
-      if (data.length >= 2048) {
-        fs.writeFileSync(SPEAKER_EMBEDDINGS_PATH, data);
-        speakerEmbeddings = new Float32Array(new Uint8Array(data).buffer);
-        return speakerEmbeddings;
-      }
-    }
-  } catch (_) {}
-  console.log('[TTS] Using bundled default speaker embedding');
-  speakerEmbeddings = defaultEmbedding();
-  const buf = Buffer.from(speakerEmbeddings.buffer);
-  fs.writeFileSync(SPEAKER_EMBEDDINGS_PATH, buf);
-  return speakerEmbeddings;
-}
-
-async function loadVoiceEmbedding(voiceId) {
-  if (!voiceId || voiceId === 'default') return ensureSpeakerEmbeddings();
-  if (voiceEmbeddingsCache.has(voiceId)) return voiceEmbeddingsCache.get(voiceId);
-  const binPath = path.join(DATA_DIR, `speaker_${voiceId}.bin`);
-  if (fs.existsSync(binPath)) {
-    const buf = fs.readFileSync(binPath);
-    const emb = new Float32Array(new Uint8Array(buf).buffer);
-    voiceEmbeddingsCache.set(voiceId, emb);
-    return emb;
-  }
-  if (voiceId.startsWith('custom_')) {
-    return generateEmbeddingFromCustomVoice(voiceId);
-  }
-  const offset = SPEAKER_OFFSETS[voiceId];
-  if (offset === undefined) return ensureSpeakerEmbeddings();
-  try {
-    const url = `${DATASET_API}&offset=${offset}&length=${SAMPLES_TO_AVERAGE}`;
-    const resp = await fetch(url);
-    if (!resp.ok) throw new Error('HTTP ' + resp.status);
-    const data = await resp.json();
-    const avg = new Float32Array(512);
-    let count = 0;
-    for (const item of data.rows) {
-      const match = item.row.filename.match(/cmu_us_(\w+)_arctic/);
-      if (match && match[1] === voiceId) {
-        for (let i = 0; i < 512; i++) avg[i] += item.row.xvector[i];
-        count++;
-      }
-    }
-    if (count === 0) return ensureSpeakerEmbeddings();
-    for (let i = 0; i < 512; i++) avg[i] /= count;
-    if (!fs.existsSync(DATA_DIR)) fs.mkdirSync(DATA_DIR, { recursive: true });
-    fs.writeFileSync(binPath, Buffer.from(avg.buffer));
-    voiceEmbeddingsCache.set(voiceId, avg);
-    return avg;
-  } catch (err) {
-    console.error('[TTS] Failed to fetch voice embedding for ' + voiceId + ':', err.message);
-    return ensureSpeakerEmbeddings();
-  }
-}
-
-let speakerFeatureExtractor = null;
-
-async function getSpeakerEmbeddingPipeline() {
-  if (speakerEmbeddingPipeline) return speakerEmbeddingPipeline;
-  if (speakerEmbeddingLoading) {
-    while (speakerEmbeddingLoading) await new Promise(r => setTimeout(r, 100));
-    if (!speakerEmbeddingPipeline) throw new Error('Speaker embedding model failed to load');
-    return speakerEmbeddingPipeline;
-  }
-  speakerEmbeddingLoading = true;
-  try {
-    const { AutoModelForXVector, AutoFeatureExtractor, env } = await loadTransformers();
-    env.allowRemoteModels = true;
-    const modelId = 'Xenova/wavlm-base-plus-sv';
-    speakerEmbeddingPipeline = await AutoModelForXVector.from_pretrained(modelId, {
-      device: 'cpu',
-      dtype: 'fp32',
-    });
-    speakerFeatureExtractor = await AutoFeatureExtractor.from_pretrained(modelId);
-    return speakerEmbeddingPipeline;
-  } catch (err) {
-    speakerEmbeddingPipeline = null;
-    speakerFeatureExtractor = null;
-    throw new Error('Speaker embedding model load failed: ' + err.message);
-  } finally {
-    speakerEmbeddingLoading = false;
-  }
-}
-
 function findCustomVoiceFile(voiceId) {
   const baseName = voiceId.replace(/^custom_/, '');
   for (const dir of getVoiceDirs()) {
@@ -278,38 +157,6 @@ async function decodeAudioFile(filePath) {
   return resampleTo16k(mono, audioBuffer.sampleRate);
 }
 
-async function generateEmbeddingFromCustomVoice(voiceId) {
-  const audioFile = findCustomVoiceFile(voiceId);
-  if (!audioFile) {
-    console.error('[VOICES] Custom voice file not found for:', voiceId);
-    return ensureSpeakerEmbeddings();
-  }
-  try {
-    console.log('[VOICES] Generating embedding from:', audioFile);
-    const audio = await decodeAudioFile(audioFile);
-    if (audio.length < SAMPLE_RATE_STT * 0.5) {
-      throw new Error('Audio too short for embedding extraction');
-    }
-    const model = await getSpeakerEmbeddingPipeline();
-    const inputs = await speakerFeatureExtractor(audio, { sampling_rate: SAMPLE_RATE_STT });
-    const output = await model(inputs);
-    const embData = output.embeddings.data;
-    const embedding = new Float32Array(512);
-    for (let i = 0; i < Math.min(512, embData.length); i++) {
-      embedding[i] = embData[i];
-    }
-    if (!fs.existsSync(DATA_DIR)) fs.mkdirSync(DATA_DIR, { recursive: true });
-    const binPath = path.join(DATA_DIR, `speaker_${voiceId}.bin`);
-    fs.writeFileSync(binPath, Buffer.from(embedding.buffer));
-    voiceEmbeddingsCache.set(voiceId, embedding);
-    console.log('[VOICES] Generated embedding for custom voice:', voiceId);
-    return embedding;
-  } catch (err) {
-    console.error('[VOICES] Failed to generate embedding for', voiceId + ':', err.message);
-    return ensureSpeakerEmbeddings();
-  }
-}
-
 async function getSTT() {
   if (sttPipeline) return sttPipeline;
   if (sttLoadError) throw sttLoadError;
@@ -342,41 +189,6 @@ async function getSTT() {
   }
 }
 
-async function getTTS() {
-  if (ttsPipeline) return ttsPipeline;
-  if (ttsLoadError) {
-    if (Date.now() - ttsLoadErrorTime < TTS_ERROR_RETRY_MS) throw ttsLoadError;
-    ttsLoadError = null;
-    ttsLoadErrorTime = 0;
-  }
-  if (ttsLoading) {
-    while (ttsLoading) await new Promise(r => setTimeout(r, 100));
-    if (ttsLoadError) throw ttsLoadError;
-    if (!ttsPipeline) throw new Error('TTS pipeline failed to load');
-    return ttsPipeline;
-  }
-  ttsLoading = true;
-  try {
-    const { pipeline, env } = await loadTransformers();
-    env.allowRemoteModels = true;
-    ttsPipeline = await pipeline('text-to-speech', 'Xenova/speecht5_tts', {
-      device: 'cpu',
-      dtype: 'fp32',
-    });
-    await ensureSpeakerEmbeddings();
-    ttsLoadError = null;
-    ttsLoadErrorTime = 0;
-    return ttsPipeline;
-  } catch (err) {
-    ttsPipeline = null;
-    ttsLoadError = new Error('TTS model load failed: ' + err.message);
-    ttsLoadErrorTime = Date.now();
-    throw ttsLoadError;
-  } finally {
-    ttsLoading = false;
-  }
-}
-
 function decodeWavToFloat32(buffer) {
   const view = new DataView(buffer.buffer || buffer);
   const riff = String.fromCharCode(view.getUint8(0), view.getUint8(1), view.getUint8(2), view.getUint8(3));
@@ -518,15 +330,11 @@ function resolveVoicePath(voiceId) {
 
 async function synthesizeViaPocket(text, voiceId) {
   const pState = pocket.getState();
-  if (!pState.healthy) return null;
-  try {
-    const voicePath = resolveVoicePath(voiceId);
-    const wav = await pocket.synthesize(text, voicePath);
-    if (wav && wav.length > 44) return wav;
-  } catch (err) {
-    console.error('[TTS] pocket-tts failed, falling back:', err.message);
-  }
-  return null;
+  if (!pState.healthy) throw new Error('pocket-tts not healthy');
+  const voicePath = resolveVoicePath(voiceId);
+  const wav = await pocket.synthesize(text, voicePath);
+  if (wav && wav.length > 44) return wav;
+  throw new Error('pocket-tts returned empty audio');
 }
 
 async function synthesize(text, voiceId) {
@@ -540,12 +348,7 @@ async function synthesize(text, voiceId) {
   const inflight = ttsInflight.get(cacheKey);
   if (inflight) return inflight;
   const promise = (async () => {
-    const pocketWav = await synthesizeViaPocket(text, voiceId);
-    if (pocketWav) { cachePut(cacheKey, pocketWav); return pocketWav; }
-    const tts = await getTTS();
-    const embeddings = await loadVoiceEmbedding(voiceId);
-    const result = await tts(text, { speaker_embeddings: embeddings });
-    const wav = encodeWav(result.audio, result.sampling_rate || SAMPLE_RATE_TTS);
+    const wav = await synthesizeViaPocket(text, voiceId);
     cachePut(cacheKey, wav);
     return wav;
   })();
@@ -555,12 +358,6 @@ async function synthesize(text, voiceId) {
 
 async function* synthesizeStream(text, voiceId) {
   const sentences = splitSentences(text);
-  const usePocket = pocket.getState().healthy;
-  let tts, embeddings;
-  if (!usePocket) {
-    tts = await getTTS();
-    embeddings = await loadVoiceEmbedding(voiceId);
-  }
   for (const sentence of sentences) {
     const cacheKey = (voiceId || 'default') + ':' + sentence;
     const cached = ttsCache.get(cacheKey);
@@ -570,28 +367,21 @@ async function* synthesizeStream(text, voiceId) {
       yield cached;
       continue;
     }
-    if (usePocket) {
-      const pocketWav = await synthesizeViaPocket(sentence, voiceId);
-      if (pocketWav) { cachePut(cacheKey, pocketWav); yield pocketWav; continue; }
-    }
-    if (!tts) { tts = await getTTS(); embeddings = await loadVoiceEmbedding(voiceId); }
-    const result = await tts(sentence, { speaker_embeddings: embeddings });
-    const wav = encodeWav(result.audio, result.sampling_rate || SAMPLE_RATE_TTS);
+    const wav = await synthesizeViaPocket(sentence, voiceId);
     cachePut(cacheKey, wav);
     yield wav;
   }
 }
 
 function getStatus() {
-  const ttsRetryExpired = ttsLoadError && (Date.now() - ttsLoadErrorTime >= TTS_ERROR_RETRY_MS);
   const pState = pocket.getState();
   return {
     sttReady: !!sttPipeline,
-    ttsReady: !!ttsPipeline || pState.healthy,
+    ttsReady: pState.healthy,
     sttLoading,
-    ttsLoading,
+    ttsLoading: false,
     sttError: sttLoadError ? sttLoadError.message : null,
-    ttsError: (ttsLoadError && !ttsRetryExpired && !pState.healthy) ? ttsLoadError.message : null,
+    ttsError: pState.healthy ? null : (pState.lastError || 'pocket-tts not running'),
     pocketTts: pState,
   };
 }
@@ -601,12 +391,9 @@ function preloadTTS() {
   const voicePath = fs.existsSync(defaultVoice) ? defaultVoice : null;
   pocket.start(voicePath).then(ok => {
     if (ok) console.log('[TTS] pocket-tts sidecar started');
-    else {
-      console.log('[TTS] pocket-tts unavailable, falling back to SpeechT5');
-      getTTS().catch(err => console.error('[TTS] SpeechT5 preload failed:', err.message));
-    }
-  }).catch(() => {
-    getTTS().catch(err => console.error('[TTS] SpeechT5 preload failed:', err.message));
+    else console.log('[TTS] pocket-tts failed to start');
+  }).catch(err => {
+    console.error('[TTS] pocket-tts start error:', err.message);
   });
 }
 
@@ -620,4 +407,4 @@ function ttsCacheGet(key) {
   return cached || null;
 }
 
-export { transcribe, synthesize, synthesizeStream, getSTT, getTTS, getStatus, getVoices, preloadTTS, ttsCacheKey, ttsCacheGet, splitSentences };
+export { transcribe, synthesize, synthesizeStream, getSTT, getStatus, getVoices, preloadTTS, ttsCacheKey, ttsCacheGet, splitSentences };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentgui",
-  "version": "1.0.179",
+  "version": "1.0.180",
   "description": "Multi-agent ACP client with real-time communication",
   "type": "module",
   "main": "server.js",