agentgui 1.0.177 → 1.0.178

package/lib/speech.js

@@ -218,24 +218,29 @@ async function loadVoiceEmbedding(voiceId) {
   }
 }
 
+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 pipeline failed to load');
+    if (!speakerEmbeddingPipeline) throw new Error('Speaker embedding model failed to load');
     return speakerEmbeddingPipeline;
   }
   speakerEmbeddingLoading = true;
   try {
-    const { pipeline, env } = await loadTransformers();
+    const { AutoModelForXVector, AutoFeatureExtractor, env } = await loadTransformers();
     env.allowRemoteModels = true;
-    speakerEmbeddingPipeline = await pipeline('feature-extraction', 'speechbrain/spkrec-xvectors-voxceleb', {
+    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;
@@ -284,11 +289,13 @@ async function generateEmbeddingFromCustomVoice(voiceId) {
     if (audio.length < SAMPLE_RATE_STT * 0.5) {
       throw new Error('Audio too short for embedding extraction');
     }
-    const pipe = await getSpeakerEmbeddingPipeline();
-    const output = await pipe(audio, { pooling: 'mean', normalize: true });
+    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, output.data.length); i++) {
-      embedding[i] = output.data[i];
+    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`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentgui",
-  "version": "1.0.177",
+  "version": "1.0.178",
   "description": "Multi-agent ACP client with real-time communication",
   "type": "module",
   "main": "server.js",