feat(stt): bessere Spracherkennung — small-Modell, VAD, feste Sprache

Gegen "Blödsinn", v. a. an Handys:
1) Modell-Default base -> small + compute_type int8 (Benchmark auf der VM:
   base patzt schon bei sauberem Audio "an Dienstag und 10 Uhr", small korrekt
   "am Dienstag um 10 Uhr"; ~4 s/6-s-Clip, gecacht).
2) faster_whisper.transcribe mit vad_filter=True (+min_silence 500ms) und
   condition_on_previous_text=False -> weniger Stille-/Kontext-Halluzinationen.
3) stt_force_language (Default True): WS-Pfad gibt die feste Zielsprache vor
   statt Auto-Detect -> robust bei kurzen/verrauschten Clips. Auto-Detect +
   Fremdsprachen-Übersetzung jetzt opt-in (stt_force_language=False).

Tests: Config-Defaults + VAD/Language-Durchreichung; test_ws-Übersetzung auf
opt-in umgestellt. Suite 253 grün.

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

app/api/ws.py

@@ -212,15 +212,21 @@ async def _voice_turn(websocket, store, user, session_id, audio, fmt, options):
         return
 
     try:
-        # Gesprochene Sprache automatisch erkennen (egal welche). Weicht sie von der
-        # Zielsprache ab, wird die Anfrage in die Zielsprache übersetzt — so erscheint
-        # und verarbeitet das System sie immer in der eingestellten Sprache.
-        transcript, detected = await orchestrator.stt.transcribe_detect(
-            audio, fmt=fmt, language=None
-        )
-        if (transcript and transcript.strip() and detected
-                and detected.lower() != (route.language or "").lower()):
-            transcript = await orchestrator.translate(transcript, route.language)
+        if settings.stt_force_language:
+            # Feste Zielsprache statt Auto-Detect -> robuster bei kurzen/verrauschten
+            # (Handy-)Clips; verhindert "falsche Sprache erkannt = Blödsinn".
+            transcript, _ = await orchestrator.stt.transcribe_detect(
+                audio, fmt=fmt, language=route.language
+            )
+        else:
+            # Gesprochene Sprache automatisch erkennen; weicht sie von der Zielsprache
+            # ab, wird in die Zielsprache übersetzt (Fremdsprachen-Diktat).
+            transcript, detected = await orchestrator.stt.transcribe_detect(
+                audio, fmt=fmt, language=None
+            )
+            if (transcript and transcript.strip() and detected
+                    and detected.lower() != (route.language or "").lower()):
+                transcript = await orchestrator.translate(transcript, route.language)
         effective_voice = voice_for_route(route.tts_provider, route.language, route.voice_gender)
     except Exception as exc:
         await websocket.send_json({"type": "error", "status": 502, "detail": str(exc)})

app/config.py

@@ -131,9 +131,10 @@ class Settings(BaseSettings):
     local_llm_max_tokens: int = 0             # 0 = serverseitiges Limit (-n)
     local_llm_temperature: float = 0.3
     local_llm_top_p: float = 0.9              # Nucleus-Sampling (0.0–1.0)
-    faster_whisper_model: str = "base"        # tiny|base|small|medium|large-v3
+    faster_whisper_model: str = "small"       # 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
+    faster_whisper_compute_type: str = "int8"     # default|int8|float16|int8_float16
+    stt_force_language: bool = True           # STT mit fester Zielsprache (robust); False = Auto-Detect + Übersetzung
     # --- Lokales TTS (piper) -------------------------------------------------
     piper_bin: str = "piper"                  # Pfad/Name des piper-Binaries
     piper_voices_dir: str = str(Path.home() / ".local" / "share" / "piper" / "voices")

app/providers/stt/faster_whisper.py

@@ -42,7 +42,13 @@ class FasterWhisperProvider(STTProvider):
         self, audio_bytes: bytes, language: str | None
     ) -> tuple[str, str | None]:
         model = _load_model(self.model_size, self.device, self.compute_type)
-        segments, info = model.transcribe(io.BytesIO(audio_bytes), language=language)
+        segments, info = model.transcribe(
+            io.BytesIO(audio_bytes),
+            language=language,
+            vad_filter=True,                                  # Stille/Rauschen entfernen -> weniger Halluzinationen
+            vad_parameters={"min_silence_duration_ms": 500},
+            condition_on_previous_text=False,                 # kurze Befehle: kein Kontext-Drift
+        )
         text = "".join(segment.text for segment in segments).strip()
         detected = getattr(info, "language", None)
         return text, detected

tests/test_stt_quality.py

@@ -0,0 +1,34 @@
+import app.providers.stt.faster_whisper as fw
+from app.config import settings
+
+
+def test_stt_config_defaults():
+    assert settings.faster_whisper_model == "small"
+    assert settings.faster_whisper_compute_type == "int8"
+    assert settings.stt_force_language is True
+
+
+class _Seg:
+    def __init__(self, text):
+        self.text = text
+
+
+class _Info:
+    language = "de"
+
+
+def test_transcribe_passes_vad_and_language(monkeypatch):
+    captured = {}
+
+    class _FakeModel:
+        def transcribe(self, audio, **kw):
+            captured.update(kw)
+            return [_Seg("hallo welt")], _Info()
+
+    monkeypatch.setattr(fw, "_load_model", lambda *a, **k: _FakeModel())
+    prov = fw.FasterWhisperProvider(model_size="small", device="cpu", compute_type="int8")
+    text, detected = prov._transcribe_sync(b"AAAA", "de")
+    assert text == "hallo welt" and detected == "de"
+    assert captured["language"] == "de"
+    assert captured["vad_filter"] is True
+    assert captured["condition_on_previous_text"] is False

tests/test_ws.py

@@ -179,6 +179,8 @@ def test_ws_voice_translates_foreign_language_to_target(monkeypatch):
     monkeypatch.setitem(deps.STT_REGISTRY, "ss", lambda s: StubSTT())
     monkeypatch.setitem(deps.LLM_REGISTRY, "ll", lambda s: StubLLM())
     monkeypatch.setitem(deps.TTS_REGISTRY, "tt", lambda s: StubTTS())
+    # Übersetzungs-Pfad ist opt-in: feste Sprache abschalten -> Auto-Detect + Übersetzung.
+    monkeypatch.setattr("app.config.settings.stt_force_language", False)
     opts = {"stt_provider": "ss", "llm_provider": "ll", "tts_provider": "tt",
             "output_endpoint": "loopback", "language": "de"}
     with client.websocket_connect("/ws/voice") as ws: