Add TTS/STT service with Coqui TTS and faster-whisper

- Added tts-stt service definition to docker-compose.yml
- Created tts-stt directory with:
  - Dockerfile: Based on debian:trixie-slim with ROCm dependencies
  - requirements.txt: Python packages including TTS, faster-whisper, FastAPI
  - app.py: FastAPI service with /tts and /stt endpoints
- Service includes GPU device mapping (/dev/kfd, /dev/dri) for ROCm acceleration
- Uses YourTTS multilingual model for TTS and large-v3 for Whisper
- Configured to use persistent storage for models and cache

ai/tts-stt/app.py

@@ -0,0 +1,128 @@
+import os
+import torch
+import numpy as np
+from TTS.api import TTS
+from faster_whisper import WhisperModel
+from fastapi import FastAPI, File, UploadFile, Form, HTTPException
+from fastapi.responses import Response
+import tempfile
+import uuid
+from pydantic import BaseModel
+from typing import Optional
+
+app = FastAPI(title="TTS-STT Service", description="Coqui TTS and faster-whisper service")
+
+# Initialize models
+tts_model = None
+whisper_model = None
+
+@app.on_event("startup")
+async def startup_event():
+    global tts_model, whisper_model
+    
+    # Get model names from environment or use defaults
+    tts_model_name = os.getenv("TTS_MODEL", "tts_models/multilingual/multi-dataset/your_tts")
+    whisper_model_name = os.getenv("WHISPER_MODEL", "large-v3")
+    device = os.getenv("DEVICE", "cuda" if torch.cuda.is_available() else "cpu")
+    
+    print(f"Loading TTS model: {tts_model_name} on {device}")
+    tts_model = TTS(model_name=tts_model_name, progress_bar=False, gpu=device=="cuda")
+    
+    print(f"Loading Whisper model: {whisper_model_name} on {device}")
+    whisper_model = WhisperModel(whisper_model_name, device=device, compute_type="float16" if device=="cuda" else "int8")
+    
+    print("Models loaded successfully")
+
+class TTSRequest(BaseModel):
+    text: str
+    language: Optional[str] = None
+    speaker_wav: Optional[str] = None  # For voice cloning with YourTTS
+
+@app.post("/tts")
+async def text_to_speech(request: TTSRequest):
+    if not tts_model:
+        raise HTTPException(status_code=503, detail="TTS model not loaded")
+    
+    try:
+        # Generate unique filename
+        filename = f"{uuid.uuid4()}.wav"
+        filepath = f"/tmp/{filename}"
+        
+        # Synthesize speech
+        if request.speaker_wav and os.path.exists(request.speaker_wav):
+            # Voice cloning
+            tts_model.tts_to_file(
+                text=request.text,
+                speaker_wav=request.speaker_wav,
+                language=request.language or "en",
+                file_path=filepath
+            )
+        else:
+            # Regular TTS
+            tts_model.tts_to_file(
+                text=request.text,
+                language=request.language or "en",
+                file_path=filepath
+            )
+        
+        # Read and return the audio file
+        with open(filepath, "rb") as f:
+            audio_data = f.read()
+        
+        # Clean up
+        os.remove(filepath)
+        
+        return Response(content=audio_data, media_type="audio/wav")
+    
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
+@app.post("/stt")
+async def speech_to_text(file: UploadFile = File(...), language: Optional[str] = None):
+    if not whisper_model:
+        raise HTTPException(status_code=503, detail="Whisper model not loaded")
+    
+    try:
+        # Save uploaded file temporarily
+        filename = f"{uuid.uuid4()}_{file.filename}"
+        filepath = f"/tmp/{filename}"
+        
+        with open(filepath, "wb") as buffer:
+            content = await file.read()
+            buffer.write(content)
+        
+        # Transcribe
+        segments, info = whisper_model.transcribe(
+            filepath,
+            language=language,
+            beam_size=5,
+            vad_filter=True,
+            vad_parameters=dict(min_silence_duration_ms=500)
+        )
+        
+        # Collect transcription
+        transcription = " ".join([segment.text for segment in segments])
+        
+        # Clean up
+        os.remove(filepath)
+        
+        return {
+            "text": transcription.strip(),
+            "language": info.language,
+            "language_probability": info.language_probability
+        }
+    
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
+@app.get("/health")
+async def health_check():
+    return {
+        "status": "healthy",
+        "tts_loaded": tts_model is not None,
+        "whisper_loaded": whisper_model is not None
+    }
+
+if __name__ == "__main__":
+    import uvicorn
+    uvicorn.run(app, host="0.0.0.0", port=8000)