feat(stt): pass conversation context as Whisper initial_prompt to reduce mishearings

src/fabledassistant/routes/voice.py

@@ -80,10 +80,12 @@ async def transcribe_audio():
         return jsonify({"error": "Audio file too large (max 25 MB)"}), 413
 
     mime_type = audio_file.content_type or "audio/webm"
+    form = await request.form
+    context = (form.get("context") or "").strip() or None
 
     t0 = time.monotonic()
     try:
-        transcript = await transcribe(audio_bytes, mime_type)
+        transcript = await transcribe(audio_bytes, mime_type, initial_prompt=context)
     except Exception:
         logger.exception("STT transcription failed")
         return jsonify({"error": "Transcription failed"}), 500

src/fabledassistant/services/stt.py

@@ -55,8 +55,12 @@ def stt_available() -> bool:
     return _model is not None
 
 
-async def transcribe(audio_bytes: bytes, mime_type: str = "audio/webm") -> str:
-    """Transcribe audio bytes to text. Runs the model in a thread executor."""
+async def transcribe(audio_bytes: bytes, mime_type: str = "audio/webm", initial_prompt: str | None = None) -> str:
+    """Transcribe audio bytes to text. Runs the model in a thread executor.
+
+    initial_prompt: optional text to bias the model toward domain-specific vocabulary
+    (e.g. recent conversation context). Reduces mishearings like "gold" for "cold".
+    """
     if _model is None:
         raise RuntimeError("STT model not loaded")
 
@@ -73,7 +77,11 @@ async def transcribe(audio_bytes: bytes, mime_type: str = "audio/webm") -> str:
             f.write(audio_bytes)
             f.flush()
             t0 = time.monotonic()
-            segments, _ = _model.transcribe(f.name, beam_size=5)  # type: ignore[union-attr]
+            segments, _ = _model.transcribe(  # type: ignore[union-attr]
+                f.name,
+                beam_size=5,
+                initial_prompt=initial_prompt or None,
+            )
             text = " ".join(seg.text.strip() for seg in segments).strip()
             logger.debug("STT transcription took %.2fs", time.monotonic() - t0)
             return text