app.py

import os
import io
import asyncio
import time
import shutil
import numpy as np
import psutil
import soundfile as sf
from concurrent.futures import ThreadPoolExecutor
from typing import Optional, Generator
from contextlib import asynccontextmanager
import logging
import aiofiles 
import torch
from fastapi import FastAPI, HTTPException, Response, StreamingResponse, UploadFile, File, Form
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel, Field

# Ensure the cloned neutts-air repository is in the path
import sys
sys.path.append(os.path.join(os.getcwd(), 'neutts-air'))
from neuttsair.neutts import NeuTTSAir

# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger("NeuTTS-API")

# --- Configuration & Utility Functions ---

# Explicitly use CPU as per Dockerfile and Hugging Face free tier compatibility
DEVICE = "cpu" 
# Configure Max Workers for concurrent synthesis threads (1-2 is safe for CPU-only)
MAX_WORKERS = 2
tts_executor = ThreadPoolExecutor(max_workers=MAX_WORKERS)
SAMPLE_RATE = 24000
CLEANUP_THRESHOLD = 3600 # 1 hour in seconds
TEMP_AUDIO_DIR = "temp_audio"
GENERATED_AUDIO_DIR = "generated_audio"
os.makedirs(TEMP_AUDIO_DIR, exist_ok=True)
os.makedirs(GENERATED_AUDIO_DIR, exist_ok=True)

class TTSRequestModel(BaseModel):
    """Model for non-file inputs to synthesis and streaming."""
    text: str = Field(..., min_length=1, max_length=1000)
    speed: float = Field(default=1.0, ge=0.5, le=2.0)
    output_format: str = Field(default="wav", pattern="^(wav|mp3|flac)$")

# --- Model Wrapper and Logic ---

class NeuTTSWrapper:
    def __init__(self, device: str = "cpu"):
        self.tts_model = None
        self.device = device
        self.load_model()

    def load_model(self):
        try:
            logger.info(f"Loading NeuTTSAir model on device: {self.device}")
            # Ensure we respect the CPU configuration
            self.tts_model = NeuTTSAir(backbone_device=self.device, codec_device=self.device)
            logger.info("✅ NeuTTSAir model loaded successfully.")
        except Exception as e:
            logger.error(f"❌ Model loading failed: {e}")
            raise

    def _convert_to_streamable_format(self, audio_data: np.ndarray, audio_format: str) -> bytes:
        """Converts NumPy audio array to streamable bytes in the specified format."""
        audio_buffer = io.BytesIO()
        try:
            sf.write(audio_buffer, audio_data, SAMPLE_RATE, format=audio_format)
        except Exception as e:
            logger.error(f"Failed to write audio data to format {audio_format}: {e}")
            raise
        audio_buffer.seek(0)
        return audio_buffer.read()

    def _split_text_into_chunks(self, text: str) -> list[str]:
        """Simple sentence splitting for streaming (can be enhanced with regex)."""
        sentences = [s.strip() for s in text.split('.') if s.strip()]
        if not sentences:
            sentences = [text.strip()]
        return sentences

    def generate_speech_blocking(self, text: str, ref_audio_path: str) -> np.ndarray:
        """Blocking synthesis for standard endpoint."""
        # 1. Load reference
        reference_audio, sr = sf.read(ref_audio_path)
        if sr != SAMPLE_RATE:
            # Simple check/resize logic required if sample rate mismatch occurs
            pass
        
        # 2. Encode reference
        ref_s = self.tts_model.encode_reference(reference_audio)
        
        # 3. Infer full text
        with torch.no_grad():
            audio = self.tts_model.infer(text, ref_s, speed=1.0)
        return audio.cpu().numpy()

    def stream_speech_blocking(self, text: str, ref_audio_path: str, speed: float, audio_format: str) -> Generator[bytes, None, None]:
        """Sentence-by-Sentence Streaming (Blocking)."""
        logger.info(f"Starting streaming synthesis for text length: {len(text)}")
        
        # 1. Load reference audio (ONLY ONCE)
        reference_audio, sr = sf.read(ref_audio_path)
        
        # 2. Encode reference (ONLY ONCE)
        ref_s = self.tts_model.encode_reference(reference_audio)
        
        # 3. Split text
        sentences = self._split_text_into_chunks(text)
        
        # 4. Stream chunks
        for i, sentence in enumerate(sentences):
            if not sentence.strip():
                continue
            
            logger.debug(f"Generating streaming chunk {i+1}: '{sentence[:30]}...'")
            
            # Infer sentence
            with torch.no_grad():
                audio_chunk = self.tts_model.infer(sentence, ref_s, speed=speed)
            
            # Convert and yield
            yield self._convert_to_streamable_format(audio_chunk.cpu().numpy(), audio_format)
            
        logger.info("Streaming synthesis complete.")

# --- Asynchronous Offloading ---

async def run_blocking_task_async(func, *args, **kwargs):
    """Offloads a blocking function call to the ThreadPoolExecutor."""
    loop = asyncio.get_event_loop()
    return await loop.run_in_executor(
        tts_executor,
        lambda: func(*args, **kwargs)
    )

async def save_upload_file_async(upload_file: UploadFile) -> str:
    """Asynchronously saves the UploadFile to disk."""
    temp_filename = os.path.join(TEMP_AUDIO_DIR, f"{time.time()}_{upload_file.filename}")
    try:
        # Use asyncio to read the file chunks in a non-blocking manner
        async with aiofiles.open(temp_filename, 'wb') as out_file:
            while content := await upload_file.read(1024 * 1024):
                await out_file.write(content)
        return temp_filename
    except Exception as e:
        logger.error(f"Error saving file: {e}")
        raise HTTPException(status_code=500, detail="Could not save reference audio file")

# --- FastAPI Lifespan Manager (Kokoro Feature) ---

@asynccontextmanager
async def lifespan(app: FastAPI):
    """Modern lifespan management: initialize model on startup, shutdown executor."""
    try:
        app.state.tts_wrapper = NeuTTSWrapper(device=DEVICE)
    except Exception as e:
        logger.error(f"Fatal startup error: {e}")
        # Terminate the application if the model can't load
        tts_executor.shutdown(wait=False) 
        raise RuntimeError("Model initialization failed.")
    
    yield # Application serves requests
    
    # Shutdown
    logger.info("Shutting down ThreadPoolExecutor.")
    tts_executor.shutdown(wait=False)

# --- FastAPI Application Setup ---
app = FastAPI(
    title="NeuTTS Air Instant Cloning API", 
    version="2.0.0-PROD-ENHANCED", 
    docs_url="/docs", 
    lifespan=lifespan
)

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_methods=["*"],
    allow_headers=["*"],
)

# --- New Endpoints and Enhancements ---

@app.get("/")
async def root():
    return {"message": "NeuTTS Air API v2.0 - Ready for Instant Voice Cloning"}

@app.get("/health")
async def health_check():
    """Enhanced health check (Kokoro Feature + Original Metrics)"""
    mem = psutil.virtual_memory()
    disk = psutil.disk_usage('/')
    
    return {
        "status": "healthy",
        "model_loaded": hasattr(app.state, 'tts_wrapper') and app.state.tts_wrapper.tts_model is not None,
        "device": DEVICE,
        "concurrency_limit": MAX_WORKERS,
        "memory_usage": {
            "total_gb": round(mem.total / (1024**3), 2),
            "used_percent": mem.percent
        },
        "disk_usage": {
            "total_gb": round(disk.total / (1024**3), 2),
            "used_percent": disk.percent
        }
    }

@app.delete("/cleanup")
async def cleanup_files():
    """Maintenance endpoint to remove old generated and temporary files."""
    await run_blocking_task_async(cleanup_files_blocking)
    return {"message": "Cleanup initiated successfully."}

def cleanup_files_blocking():
    """Blocking file cleanup logic (original NeuTTS feature)."""
    now = time.time()
    deleted_count = 0
    
    for directory in [GENERATED_AUDIO_DIR, TEMP_AUDIO_DIR]:
        for filename in os.listdir(directory):
            filepath = os.path.join(directory, filename)
            if os.path.isfile(filepath):
                try:
                    # Original cleanup logic: delete if older than CLEANUP_THRESHOLD
                    if now - os.path.getctime(filepath) > CLEANUP_THRESHOLD:
                        os.remove(filepath)
                        deleted_count += 1
                except Exception as e:
                    logger.warning(f"Failed to delete {filepath}: {e}")
    
    logger.info(f"Cleanup completed: {deleted_count} files removed.")
    return deleted_count


# --- Core Synthesis Endpoints ---

@app.post("/synthesize", response_class=Response)
async def text_to_speech(
    text: str = Form(...),
    speed: float = Form(1.0, ge=0.5, le=2.0),
    output_format: str = Form("wav", pattern="^(wav|mp3|flac)$"),
    reference_audio: UploadFile = File(...)
):
    """
    Standard blocking TTS endpoint with Multi-Format Output (Kokoro Feature).
    Uses ThreadPoolExecutor for non-blocking API responsiveness.
    """
    if not hasattr(app.state, 'tts_wrapper'):
        raise HTTPException(status_code=503, detail="Service unavailable: Model not loaded")

    # 1. Asynchronously save reference audio
    temp_ref_path = await save_upload_file_async(reference_audio)
    start_time = time.time()
    
    try:
        # 2. Offload the ENTIRE blocking process (encode + infer) to a thread
        audio_data = await run_blocking_task_async(
            app.state.tts_wrapper.generate_speech_blocking,
            text,
            temp_ref_path
        )
        
        # 3. Convert to requested format (Blocking, but usually fast)
        audio_bytes = await run_blocking_task_async(
            app.state.tts_wrapper._convert_to_streamable_format,
            audio_data,
            output_format
        )
        
        # 4. Save to disk (Original NeuTTS requirement)
        audio_filename = f"tts_{time.time()}.{output_format}"
        final_path = os.path.join(GENERATED_AUDIO_DIR, audio_filename)
        # We perform the file write operation in a blocking manner inside the thread pool.
        await run_blocking_task_async(
            lambda: open(final_path, 'wb').write(audio_bytes)
        )
        
        processing_time = time.time() - start_time
        audio_duration = len(audio_data) / SAMPLE_RATE

        return Response(
            content=audio_bytes,
            media_type=f"audio/{'mpeg' if output_format == 'mp3' else output_format}",
            headers={
                "Content-Disposition": f"attachment; filename={audio_filename}",
                "X-Processing-Time": f"{processing_time:.2f}s",
                "X-Audio-Duration": f"{audio_duration:.2f}s"
            }
        )

    except Exception as e:
        logger.error(f"Synthesis error: {e}")
        raise HTTPException(status_code=500, detail=f"Synthesis failed: {e}")
    finally:
        # 5. Clean up the temporary reference file
        if os.path.exists(temp_ref_path):
            os.unlink(temp_ref_path)

@app.post("/synthesize/stream")
async def stream_text_to_speech_cloning(
    text: str = Form(..., min_length=1, max_length=5000), # Increased limit for streaming
    speed: float = Form(1.0, ge=0.5, le=2.0),
    output_format: str = Form("mp3", pattern="^(wav|mp3|flac)$"), # MP3 is best for streaming
    reference_audio: UploadFile = File(...)
):
    """
    Sentence-by-Sentence Streaming Endpoint (Kokoro Feature adaptation).
    Performs encoding once, then synthesizes and streams chunks.
    """
    if not hasattr(app.state, 'tts_wrapper'):
        raise HTTPException(status_code=503, detail="Service unavailable: Model not loaded")

    # 1. Asynchronously save reference audio (non-blocking)
    temp_ref_path = await save_upload_file_async(reference_audio)
    
    # 2. Define the generator function, which will run in the thread pool implicitly
    def stream_generator():
        try:
            # The entire streaming process runs blocking inside the thread pool
            for chunk_bytes in app.state.tts_wrapper.stream_speech_blocking(
                text, 
                temp_ref_path, 
                speed,
                output_format
            ):
                yield chunk_bytes
        except Exception as e:
            logger.error(f"Streaming generator error: {e}")
            # Raise an exception if necessary, though it might break the stream
        finally:
            # 3. Cleanup the temporary reference file after the stream is done
            if os.path.exists(temp_ref_path):
                os.unlink(temp_ref_path)
            
    # The StreamingResponse handles the transfer encoding and chunking
    return StreamingResponse(
        stream_generator(),
        media_type=f"audio/{'mpeg' if output_format == 'mp3' else output_format}",
        headers={
            "Content-Disposition": "attachment; filename=tts_live_stream.mp3",
            "Transfer-Encoding": "chunked",
            "Cache-Control": "no-cache"
        }
    )

@app.get("/audio/{filename}")
async def get_audio(filename: str):
    """Original NeuTTS feature to serve generated audio files."""
    file_path = os.path.join(GENERATED_AUDIO_DIR, filename)
    if not os.path.exists(file_path):
        raise HTTPException(status_code=404, detail="Audio file not found")
    
    return Response(
        content=open(file_path, "rb").read(),
        media_type=f"audio/{filename.split('.')[-1]}", # Simple media type detection
        headers={"Content-Disposition": f"attachment; filename={filename}"}
    )