Feat: Edge-TTS-Engine (native dt. Azure-Stimmen, kein Akzent) + Probe-hören-Knopf

Edge-TTS als 4. Engine (gratis, kein Key, KEIN Cloning → natives Deutsch ohne Akzent — die einzige
Lösung gegen das Akzent-Problem aller Cloning-Engines). /voices listet dt. Edge-Stimmen (weiblich
zuerst, inkl. Gisela). Picker: Engine 'Edge (natürlich · gratis)' + Probe-hören-Knopf (festen Satz je
Engine/Stimme abspielen, ohne reinsprechen). Default bleibt Piper.

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

voice_service/app.py

@@ -60,6 +60,8 @@ CHATTERBOX_REF = os.environ.get("VOICE_CHATTERBOX_REF", "")
 ELEVEN_MODEL = os.environ.get("VOICE_ELEVENLABS_MODEL", "eleven_flash_v2_5")
 ELEVEN_DEFAULT_VOICE = os.environ.get("VOICE_ELEVENLABS_VOICE", "21m00Tcm4TlvDq8ikWAM")
 HERMES_ENV = Path.home() / ".hermes" / ".env"
+# Edge-TTS: native deutsche Azure-Neural-Stimmen, gratis & ohne Key (kein Cloning → kein Akzent).
+EDGE_DEFAULT = os.environ.get("VOICE_EDGE_VOICE", "de-DE-KatjaNeural")
 
 
 def _eleven_key() -> str:
@@ -230,6 +232,60 @@ def elevenlabs_list() -> list[dict]:
         return []
 
 
+# =================================================================================
+# TTS — Edge-TTS (native dt. Azure-Neural-Stimmen, gratis, kein Key, KEIN Cloning → kein Akzent)
+# =================================================================================
+def edge_tts_synth(text: str, voice: str) -> bytes:
+    import asyncio
+    import edge_tts
+
+    v = voice or EDGE_DEFAULT
+
+    async def _run() -> bytes:
+        buf = bytearray()
+        communicate = edge_tts.Communicate(text, v)
+        async for chunk in communicate.stream():
+            if chunk["type"] == "audio":
+                buf.extend(chunk["data"])
+        return bytes(buf)
+
+    try:
+        return asyncio.run(_run())
+    except Exception as exc:  # noqa: BLE001
+        raise HTTPException(502, f"Edge-TTS-Fehler: {exc}")
+
+
+_edge_cache: list | None = None
+
+
+def edge_list() -> list[dict]:
+    """Deutsche Edge/Azure-Stimmen (einmal vom Dienst geholt + gecacht). Weibliche zuerst."""
+    global _edge_cache
+    if _edge_cache is not None:
+        return _edge_cache
+    try:
+        import asyncio
+        import edge_tts
+
+        voices = asyncio.run(edge_tts.list_voices())
+        out = []
+        for v in voices:
+            if not v.get("Locale", "").startswith("de-"):
+                continue
+            short = v["ShortName"]                       # z.B. de-DE-KatjaNeural
+            gender = "weiblich" if v.get("Gender") == "Female" else "männlich"
+            name = short.split("-")[-1].replace("Neural", "").replace("Multilingual", " (multiling.)")
+            out.append({"engine": "edge", "id": short,
+                        "label": f"{name} · {v['Locale']} · {gender}",
+                        "clonable": False, "_female": v.get("Gender") == "Female"})
+        out.sort(key=lambda i: (not i.pop("_female"), i["id"]))
+        _edge_cache = out
+    except Exception:  # noqa: BLE001
+        log.warning("Edge-Stimmenliste fehlgeschlagen", exc_info=True)
+        _edge_cache = []
+    return _edge_cache
+
+
 # =================================================================================
 # FastAPI
 # =================================================================================
@@ -258,14 +314,14 @@ class TTSIn(BaseModel):
 @app.get("/health")
 def health() -> dict:
     return {"ok": True, "stt_model": STT_MODEL,
-            "engines": ["piper", "chatterbox", "elevenlabs"],
+            "engines": ["edge", "piper", "chatterbox", "elevenlabs"],
             "elevenlabs_key": bool(_eleven_key()),
             "piper_voices": [v["id"] for v in piper_list()]}
 
 
 @app.get("/voices")
 def voices() -> dict:
-    return {"voices": piper_list() + chatterbox_list() + elevenlabs_list(),
+    return {"voices": edge_list() + piper_list() + chatterbox_list() + elevenlabs_list(),
             "default": {"engine": "piper", "voice": PIPER_DEFAULT}}
 
 
@@ -284,6 +340,8 @@ def tts(body: TTSIn) -> Response:
     text = (body.text or "").strip()
     if not text:
         raise HTTPException(400, "Leerer Text.")
+    if body.engine == "edge":
+        return Response(content=edge_tts_synth(text, body.voice), media_type="audio/mpeg")
     if body.engine == "elevenlabs":
         return Response(content=elevenlabs_tts(text, body.voice), media_type="audio/mpeg")
     if body.engine == "chatterbox":