feat(stt): echtes lokales STT via faster-whisper (optional .[local])

- FasterWhisperProvider implementiert (CTranslate2): Modell prozessweit gecacht
  (lru_cache), Transkription in asyncio.to_thread; robuster CPU-Fallback wenn
  GPU/compute_type nicht verfuegbar
- Config: FASTER_WHISPER_MODEL/DEVICE/COMPUTE_TYPE (Defaults base/auto/default)
- pyproject: optionales Extra [local] = faster-whisper
- Test: transcribe-Endpunkt nutzt jetzt einen Stub-STT (kein Platzhalter mehr)
- Doku: Voraussetzungen, Hybrid-Beispiel (STT+LLM lokal, TTS remote), Architektur/README

Live verifiziert: TTS->WAV->faster-whisper transkribiert korrekt (de); voller Hybrid
ueber /ws/voice (faster-whisper + Ollama llama3.2 + OpenRouter-TTS) funktioniert.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app/config.py

@@ -120,6 +120,9 @@ class Settings(BaseSettings):
     local_llm_base_url: str = "http://127.0.0.1:11434/v1"
     local_llm_api_key: str = "dummy"
     local_llm_model: str = "llama3.1"
+    faster_whisper_model: str = "base"        # tiny|base|small|medium|large-v3
+    faster_whisper_device: str = "auto"       # auto|cpu|cuda
+    faster_whisper_compute_type: str = "default"  # default|int8|float16|int8_float16
     db_path: str = str(BASE_DIR / "data" / "voice-assistant.db")
     admin_api_key: str = ""
     auth_enabled: bool = True

app/providers/stt/faster_whisper.py

@@ -1,5 +1,49 @@
+"""Lokaler STT-Provider auf Basis von faster-whisper (CTranslate2).
+
+Optionale Dependency: `pip install -e .[local]`. Das Whisper-Modell wird beim
+ersten Aufruf geladen (und ggf. heruntergeladen) und prozessweit zwischengespeichert.
+Die Transkription ist CPU/GPU-lastig und laeuft daher in einem Thread, damit der
+Event-Loop frei bleibt.
+"""
+
+import asyncio
+import io
+from functools import lru_cache
+
+from app.config import settings
 from app.providers.stt.base import STTProvider
 
+
+@lru_cache(maxsize=2)
+def _load_model(model_size: str, device: str, compute_type: str):
+    try:
+        from faster_whisper import WhisperModel
+    except ModuleNotFoundError as exc:  # pragma: no cover - haengt von Installation ab
+        raise RuntimeError(
+            "faster-whisper ist nicht installiert. Installieren mit: pip install -e .[local]"
+        ) from exc
+    try:
+        return WhisperModel(model_size, device=device, compute_type=compute_type)
+    except Exception:
+        # GPU/Compute-Type nicht verfuegbar -> robuster CPU-Fallback (int8).
+        if device != "cpu":
+            return WhisperModel(model_size, device="cpu", compute_type="int8")
+        raise
+
+
 class FasterWhisperProvider(STTProvider):
+    def __init__(self, model_size: str | None = None, device: str | None = None,
+                 compute_type: str | None = None):
+        self.model_size = model_size or settings.faster_whisper_model
+        self.device = device or settings.faster_whisper_device
+        self.compute_type = compute_type or settings.faster_whisper_compute_type
+
+    def _transcribe_sync(self, audio_bytes: bytes, language: str | None) -> str:
+        model = _load_model(self.model_size, self.device, self.compute_type)
+        segments, _info = model.transcribe(io.BytesIO(audio_bytes), language=language)
+        return "".join(segment.text for segment in segments).strip()
+
     async def transcribe(self, audio_bytes: bytes, fmt: str, language: str | None = None) -> str:
-        return "[local transcription placeholder]"
+        if not audio_bytes:
+            raise ValueError("STT input audio is empty")
+        return await asyncio.to_thread(self._transcribe_sync, audio_bytes, language)