server.py

import json
import os.path
import tempfile
import sys
import re
import uuid
import requests
import librosa
import numpy as np
import torch
import uvicorn
import torchaudio
import base64
import io
from argparse import ArgumentParser
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from transformers import WhisperFeatureExtractor, AutoTokenizer
from speech_tokenizer.modeling_whisper import WhisperVQEncoder

sys.path.insert(0, "./cosyvoice")
sys.path.insert(0, "./third_party/Matcha-TTS")

from speech_tokenizer.utils import extract_speech_token
from flow_inference import AudioDecoder

# Initialize FastAPI
app = FastAPI()
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

class AudioRequest(BaseModel):
    audio_data: str  # Base64 encoded audio
    sample_rate: int

class AudioResponse(BaseModel):
    audio_data: str  # Base64 encoded response audio
    text_transcript: str

# Global initialization
DEVICE = "cuda"
audio_decoder = None
whisper_model = None
feature_extractor = None
glm_tokenizer = None

def initialize_models():
    global audio_decoder, feature_extractor, whisper_model, glm_tokenizer
    
    # Initialize Whisper and GLM components
    glm_tokenizer = AutoTokenizer.from_pretrained("THUDM/glm-4-voice-9b", trust_remote_code=True)
    
    # Initialize Whisper model
    whisper_model = WhisperVQEncoder.from_pretrained("THUDM/glm-4-voice-tokenizer").eval().to(DEVICE)
    feature_extractor = WhisperFeatureExtractor.from_pretrained("THUDM/glm-4-voice-tokenizer")
    
    # Initialize AudioDecoder
    audio_decoder = AudioDecoder(
        config_path="./glm-4-voice-decoder/config.yaml",
        flow_ckpt_path="./glm-4-voice-decoder/flow.pt",
        hift_ckpt_path="./glm-4-voice-decoder/hift.pt",
        device=DEVICE
    )

@app.on_event("startup")
async def startup_event():
    try:
        initialize_models()
    except Exception as e:
        raise HTTPException(status_code=500, detail=f"Model initialization failed: {str(e)}")

def process_audio(audio_bytes: bytes, target_sr: int = 16000):
    # Convert bytes to numpy array
    audio_np = np.frombuffer(audio_bytes, dtype=np.int16)
    
    # Resample if necessary
    if target_sr != 16000:  # Whisper's default sample rate
        audio_np = librosa.resample(audio_np, orig_sr=target_sr, target_sr=16000)
    
    return audio_np

@app.post("/api/voice_chat")
async def voice_chat(request: AudioRequest):
    try:
        # Decode and process audio
        audio_bytes = base64.b64decode(request.audio_data)
        audio_np = process_audio(audio_bytes, request.sample_rate)
        
        # Extract tokens
        with tempfile.TemporaryDirectory() as tmp_dir:
            tmp_path = os.path.join(tmp_dir, "audio.wav")
            torchaudio.save(tmp_path, torch.from_numpy(audio_np).unsqueeze(0), request.sample_rate)
            audio_tokens = extract_speech_token(whisper_model, feature_extractor, [tmp_path])[0]

        if not audio_tokens:
            raise HTTPException(400, "No speech detected")

        # Generate response
        response = requests.post(
            "http://localhost:10000/generate_stream",
            json={
                "prompt": f"<|system|>Respond<|user|>{' '.join(f'<|audio_{x}|>' for x in audio_tokens)}<|assistant|>",
                "temperature": 0.7,
                "top_p": 0.9,
                "max_new_tokens": 256
            },
            stream=True
        )

        # Process response
        text_tokens = []
        audio_tokens = []
        audio_offset = glm_tokenizer.convert_tokens_to_ids('<|audio_0|>')
        
        for chunk in response.iter_lines():
            token_id = json.loads(chunk)["token_id"]
            if token_id >= audio_offset:
                audio_tokens.append(token_id - audio_offset)
            else:
                text_tokens.append(token_id)

        # Generate audio
        tts_token = torch.tensor(audio_tokens, device=DEVICE).unsqueeze(0)
        tts_speech, _ = audio_decoder.token2wav(tts_token)
        
        # Prepare response
        buffer = io.BytesIO()
        torchaudio.save(buffer, tts_speech.cpu(), 22050, format="wav")
        
        return AudioResponse(
            audio_data=base64.b64encode(buffer.getvalue()).decode(),
            text_transcript=glm_tokenizer.decode(text_tokens, skip_special_tokens=True)
        )
        
    except Exception as e:
        raise HTTPException(500, str(e))
    
if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000)