Fix speech gate, hallucination filter, and latency

Speech gate (silence detection):
- vadHadSpeech flag: VAD auto-stop cancels without calling STT when no
  speech was detected (fixes "[STT] No speech detected → gibberish" loop)
- cancelNextBlob flag: onstop skips processBlob when VAD cancels silently
- vadLastVoiceMs: gates preview transcription on actual detected speech
  (prevents "reich" hallucination on initial silence chunks)

Hallucination filter:
- Client: HALLUCINATION_RE strips known Whisper phantoms from preview
- Server: _is_hallucination() in generate() treats "reich" / "danke" /
  "thank you" etc. as "No speech detected" → never reaches LLM

Latency:
- VAD_SILENCE_MS: 1500 → 1000 ms (sends 500 ms sooner per turn)
- VAD_MIN_REC_MS: 500 → 400 ms
- MediaRecorder timeslice: 2500 → 1500 ms (preview text updates faster)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

routes/conversation.py

@@ -33,6 +33,19 @@ from routes.stt import _transcribe_audio, _clean_stt_backend
 
 router = APIRouter()
 
+# ── STT hallucination filter ──────────────────────────────────────────────────
+# Whisper commonly hallucinates these phrases on silence/noise.
+# Treat them as "no speech detected" rather than passing them to the LLM.
+_HALLUCINATIONS: frozenset[str] = frozenset([
+    "reich", "danke", "danke schön", "danke schoen", "vielen dank",
+    "thank you", "thank you.", "thanks", "thanks.", "you", "you.",
+    "copyright", "abonnieren", "untertitel", "subscribe", "subscribing",
+])
+
+def _is_hallucination(text: str) -> bool:
+    t = text.strip().lower().rstrip(".!?,;:-").strip()
+    return len(t) <= 2 or t in _HALLUCINATIONS
+
 # ── Sentence-boundary helpers for pipelined TTS ───────────────────────────────
 
 _SENT_RE = re.compile(r'(?<=[.!?])\s+')
@@ -749,7 +762,7 @@ async def conversation_turn(
                     with contextlib.suppress(Exception):
                         p.unlink(missing_ok=True)
 
-        if not transcript.strip():
+        if not transcript.strip() or _is_hallucination(transcript):
             yield sse({"type": "error", "stage": "stt", "message": "No speech detected."})
             return
 

static/js/conversation.js

@@ -178,8 +178,10 @@ $('s-import-voices-file')?.addEventListener('change', async function () {
   let liveInterimText = '';
   let speechRec = null;
   const VAD_THRESHOLD = 0.01;
-  const VAD_MIN_REC_MS = 500;
+  const VAD_MIN_REC_MS = 400;
-  const VAD_SILENCE_MS = 1500;
+  const VAD_SILENCE_MS = 1000;
+  // Whisper hallucinations on silence/noise — discard these from the live preview
+  const HALLUCINATION_RE = /^(reich|danke\s*(schön)?|vielen\s*dank|thank\s*you|thanks|you|copyright|abonnieren|untertitel|zарегистрируйтесь)[.!?,\s]*$/i;
   const origPlaceholder = textInput?.placeholder || '';
   const vadToggle       = $('conv-vad-toggle');
   const handsFreeToggle = $('conv-handsfree-toggle');
@@ -189,6 +191,9 @@ $('s-import-voices-file')?.addEventListener('change', async function () {
   const audioQueue = [];
   let audioQueuePlaying = false;
   let audioQueueDrainCb = null;
+  let vadHadSpeech = false;    // true once RMS crossed threshold during this recording
+  let vadLastVoiceMs = 0;      // last timestamp speech was detected (for preview gate)
+  let cancelNextBlob = false;  // set by VAD when no speech was detected → skip STT
 
   // ── Populate STT backends ────────────────────────────────────────────────
   async function loadConvSttBackends() {
@@ -421,6 +426,8 @@ $('s-import-voices-file')?.addEventListener('change', async function () {
   // ── Chunked Whisper preview (fallback for browsers without SpeechRecognition) ──
   async function transcribeForPreview() {
     if (previewTranscribing || !recChunks.length) return;
+    // Only transcribe if speech was actually detected in this recording
+    if (!vadHadSpeech || Date.now() - vadLastVoiceMs > 5000) return;
     previewTranscribing = true;
     try {
       const mime = (mediaRecorder && mediaRecorder.mimeType) || 'audio/webm';
@@ -433,7 +440,8 @@ $('s-import-voices-file')?.addEventListener('change', async function () {
       if (r.ok) {
         const d = await r.json();
         const txt = (d.text || '').trim();
-        if (txt && mediaRecorder && mediaRecorder.state === 'recording') {
+        if (txt && txt.length > 3 && !HALLUCINATION_RE.test(txt)
+            && mediaRecorder && mediaRecorder.state === 'recording') {
           liveInterimText = txt;
           if (textInput) textInput.value = txt;
         }
@@ -477,22 +485,36 @@ $('s-import-voices-file')?.addEventListener('change', async function () {
 
     liveInterimText = '';
     previewTranscribing = false;
+    vadHadSpeech = false;
+    vadLastVoiceMs = 0;
+    cancelNextBlob = false;
     recChunks = [];
     mediaRecorder = new MediaRecorder(stream);
     mediaRecorder.ondataavailable = e => {
       if (e.data.size > 0) {
         recChunks.push(e.data);
-        // Trigger Whisper preview on periodic chunks (not on the final stop chunk)
+        // Only run preview when speech was detected (prevents "reich" on silence chunks)
-        if (mediaRecorder.state === 'recording') transcribeForPreview();
+        if (mediaRecorder.state === 'recording' && vadHadSpeech) transcribeForPreview();
       }
     };
     mediaRecorder.onstop = () => {
       stream.getTracks().forEach(t => t.stop());
+      if (cancelNextBlob) {
+        // VAD fired but no speech was detected — reset without calling STT
+        cancelNextBlob = false;
+        if (textInput) { textInput.readOnly = false; textInput.classList.remove('listening'); textInput.placeholder = origPlaceholder; textInput.value = ''; }
+        if (sendBtn) sendBtn.disabled = false;
+        micBtn.classList.remove('processing');
+        micIcon.className = 'mdi mdi-microphone';
+        if (micStatus) micStatus.textContent = 'Ready';
+        isProcessing = false;
+        return;
+      }
       const blob = new Blob(recChunks, { type: mediaRecorder.mimeType || 'audio/webm' });
       processBlob(blob);
     };
-    // timeslice=2500: ondataavailable fires every 2.5 s → interim Whisper preview
+    // timeslice=1500: ondataavailable fires every 1.5 s → faster interim Whisper preview
-    mediaRecorder.start(2500);
+    mediaRecorder.start(1500);
     micBtn.classList.add('recording');
     micIcon.className = 'mdi mdi-stop';
     if (micStatus) micStatus.textContent = 'Recording…';
@@ -549,7 +571,11 @@ $('s-import-voices-file')?.addEventListener('change', async function () {
           for (const s of vadBuf) rms += s * s;
           rms = Math.sqrt(rms / vadBuf.length);
           if (levelFill) levelFill.style.width = Math.min(100, rms * 5000) + '%';
-          if (rms > VAD_THRESHOLD) vadLastVoice = Date.now();
+          if (rms > VAD_THRESHOLD) {
+            vadLastVoice = Date.now();
+            vadLastVoiceMs = Date.now();
+            vadHadSpeech = true;
+          }
           const elapsed = Date.now() - recStart;
           const silence = Date.now() - vadLastVoice;
           if (elapsed > VAD_MIN_REC_MS) {
@@ -560,6 +586,11 @@ $('s-import-voices-file')?.addEventListener('change', async function () {
                 : 'Recording…';
             }
             if (silence >= VAD_SILENCE_MS) {
+              if (!vadHadSpeech) {
+                // Silence the whole time — cancel without calling STT
+                cancelNextBlob = true;
+                if (micStatus) micStatus.textContent = 'Ready';
+              }
               stopRecording();
               return;
             }