feat(tts): add Edge TTS (Microsoft Neural Voices) as primary engine with Piper fallback

Edge TTS provides near-human quality voices (de-DE-ConradNeural, en-US-GuyNeural).
Falls back to Piper TTS when Edge TTS is unavailable (e.g. no internet).

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

compliance-tts-service/Dockerfile

@@ -15,8 +15,16 @@ WORKDIR /app
 COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 
-# Download Piper voice model (German, thorsten, high quality)
+# Download Piper voice models
-RUN mkdir -p /app/models &&     wget -q -O /app/models/de_DE-thorsten-high.onnx     "https://huggingface.co/rhasspy/piper-voices/resolve/main/de/de_DE/thorsten/high/de_DE-thorsten-high.onnx" &&     wget -q -O /app/models/de_DE-thorsten-high.onnx.json     "https://huggingface.co/rhasspy/piper-voices/resolve/main/de/de_DE/thorsten/high/de_DE-thorsten-high.onnx.json"
+RUN mkdir -p /app/models && \
+    wget -q -O /app/models/de_DE-thorsten-high.onnx \
+    "https://huggingface.co/rhasspy/piper-voices/resolve/main/de/de_DE/thorsten/high/de_DE-thorsten-high.onnx" && \
+    wget -q -O /app/models/de_DE-thorsten-high.onnx.json \
+    "https://huggingface.co/rhasspy/piper-voices/resolve/main/de/de_DE/thorsten/high/de_DE-thorsten-high.onnx.json" && \
+    wget -q -O /app/models/en_US-lessac-high.onnx \
+    "https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/lessac/high/en_US-lessac-high.onnx" && \
+    wget -q -O /app/models/en_US-lessac-high.onnx.json \
+    "https://huggingface.co/rhasspy/piper-voices/resolve/main/en/en_US/lessac/high/en_US-lessac-high.onnx.json"
 
 # Copy application
 COPY . .

compliance-tts-service/main.py

@@ -1,10 +1,12 @@
 """Compliance TTS Service — Piper TTS + FFmpeg Audio/Video Pipeline."""
+import hashlib
 import logging
 import os
 import tempfile
 import uuid
 
 from fastapi import FastAPI, HTTPException
+from fastapi.responses import FileResponse, Response
 from pydantic import BaseModel
 
 from storage import StorageClient
@@ -21,6 +23,7 @@ MINIO_ACCESS_KEY = os.getenv("MINIO_ACCESS_KEY", "breakpilot")
 MINIO_SECRET_KEY = os.getenv("MINIO_SECRET_KEY", "breakpilot123")
 MINIO_SECURE = os.getenv("MINIO_SECURE", "false").lower() == "true"
 PIPER_MODEL_PATH = os.getenv("PIPER_MODEL_PATH", "/app/models/de_DE-thorsten-high.onnx")
+PIPER_MODEL_EN_PATH = os.getenv("PIPER_MODEL_EN_PATH", "/app/models/en_US-lessac-high.onnx")
 
 AUDIO_BUCKET = "compliance-training-audio"
 VIDEO_BUCKET = "compliance-training-video"
@@ -28,6 +31,7 @@ VIDEO_BUCKET = "compliance-training-video"
 # Initialize services
 storage = StorageClient(MINIO_ENDPOINT, MINIO_ACCESS_KEY, MINIO_SECRET_KEY, secure=MINIO_SECURE)
 tts = PiperTTS(PIPER_MODEL_PATH)
+tts_en = PiperTTS(PIPER_MODEL_EN_PATH) if os.path.exists(PIPER_MODEL_EN_PATH) else None
 
 
 @app.on_event("startup")
@@ -104,16 +108,108 @@ async def health():
 @app.get("/voices")
 async def list_voices():
     """List available TTS voices."""
-    return {
+    voices = [
-        "voices": [
+        VoiceInfo(
-            VoiceInfo(
+            id="de_DE-thorsten-high",
-                id="de_DE-thorsten-high",
+            language="de",
-                language="de",
+            name="Thorsten (High Quality)",
-                name="Thorsten (High Quality)",
+            quality="high",
-                quality="high",
+        ),
-            ),
+    ]
-        ],
+    if tts_en is not None:
-    }
+        voices.append(VoiceInfo(
+            id="en_US-lessac-high",
+            language="en",
+            name="Lessac (High Quality)",
+            quality="high",
+        ))
+    return {"voices": voices}
+
+
+class SynthesizeDirectRequest(BaseModel):
+    text: str
+    language: str = "de"
+
+
+# Simple disk cache for synthesized audio (avoids re-synthesis of same text)
+TTS_CACHE_DIR = "/tmp/tts-cache"
+os.makedirs(TTS_CACHE_DIR, exist_ok=True)
+
+
+EDGE_TTS_VOICES = {
+    "de": "de-DE-ConradNeural",
+    "en": "en-US-GuyNeural",
+}
+
+
+async def _edge_tts_synthesize(text: str, language: str, output_path: str) -> bool:
+    """Synthesize using Edge TTS (Microsoft Neural Voices). Returns True on success."""
+    try:
+        import edge_tts
+        voice = EDGE_TTS_VOICES.get(language, EDGE_TTS_VOICES["de"])
+        communicate = edge_tts.Communicate(text, voice)
+        await communicate.save(output_path)
+        return True
+    except Exception as e:
+        logger.warning(f"Edge TTS failed, falling back to Piper: {e}")
+        return False
+
+
+@app.post("/synthesize-direct")
+async def synthesize_direct(req: SynthesizeDirectRequest):
+    """Synthesize text and return MP3 audio directly (no MinIO upload).
+
+    Uses Edge TTS (Microsoft Neural Voices) for high-quality speech.
+    Falls back to Piper TTS if Edge TTS is unavailable (e.g. no internet).
+    Includes disk caching so identical text is only synthesized once.
+    """
+    if not req.text.strip():
+        raise HTTPException(status_code=400, detail="Text is empty")
+
+    # Cache key based on text + language hash
+    text_hash = hashlib.sha256(f"{req.language}:{req.text}".encode()).hexdigest()[:16]
+    cache_path = os.path.join(TTS_CACHE_DIR, f"{text_hash}.mp3")
+
+    if os.path.exists(cache_path):
+        logger.info(f"TTS cache hit: {text_hash}")
+        return FileResponse(
+            cache_path,
+            media_type="audio/mpeg",
+            headers={"X-TTS-Cache": "hit"},
+        )
+
+    # Try Edge TTS first (high quality neural voices)
+    success = await _edge_tts_synthesize(req.text, req.language, cache_path)
+
+    if success and os.path.exists(cache_path):
+        size = os.path.getsize(cache_path)
+        logger.info(f"Edge TTS ({req.language}): {len(req.text)} chars, {size} bytes, cached as {text_hash}")
+        return FileResponse(
+            cache_path,
+            media_type="audio/mpeg",
+            headers={"X-TTS-Cache": "miss", "X-TTS-Engine": "edge"},
+        )
+
+    # Fallback: Piper TTS
+    engine = tts
+    if req.language == "en" and tts_en is not None:
+        engine = tts_en
+
+    with tempfile.TemporaryDirectory() as tmpdir:
+        try:
+            mp3_path, duration = engine.synthesize_to_mp3(req.text, tmpdir)
+            import shutil
+            shutil.copy2(mp3_path, cache_path)
+            logger.info(f"Piper TTS ({req.language}): {len(req.text)} chars, {duration:.1f}s, cached as {text_hash}")
+        except Exception as e:
+            logger.error(f"TTS synthesis failed: {e}")
+            raise HTTPException(status_code=500, detail=str(e))
+
+    return FileResponse(
+        cache_path,
+        media_type="audio/mpeg",
+        headers={"X-TTS-Cache": "miss", "X-TTS-Engine": "piper"},
+    )
 
 
 @app.post("/presigned-url", response_model=PresignedURLResponse)

compliance-tts-service/requirements.txt

@@ -3,3 +3,4 @@ uvicorn[standard]==0.27.1
 boto3==1.34.25
 python-multipart==0.0.6
 pydantic==2.6.1
+edge-tts==6.1.12