tts-server/speech_tts_cpu.py

from fastapi import FastAPI, HTTPException, Body, Request
from fastapi.responses import StreamingResponse
from contextlib import asynccontextmanager
from kokoro import KPipeline
import soundfile as sf
import io
import base64
import re
import logging
import json
import hashlib
import asyncio
import os
import uuid
import aiofiles
from fastapi.middleware.cors import CORSMiddleware
from typing import Dict, List, Optional
from cachetools import TTLCache
import concurrent.futures
import threading
import numpy as np
from pathlib import Path
import sys
os.environ["USE_NNPACK"] = "0"
os.environ["NNPACK_DISABLE"] = "1"
os.environ["TORCH_CPP_LOG_LEVEL"] = "ERROR" 
import torch
torch.backends.nnpack.enabled = False
import warnings
warnings.filterwarnings("ignore", message=".*NNPACK.*")

# ------------------- 配置 -------------------
USE_MP3 = os.getenv("USE_MP3", "false").lower() == "true"
MEMORY_CACHE_SIZE = int(os.getenv("MEMORY_CACHE_SIZE", 200))
DISK_CACHE_SIZE = int(os.getenv("DISK_CACHE_SIZE", 500))
MAX_CONCURRENT_REQUESTS = int(os.getenv("MAX_CONCURRENT_REQUESTS", 12))
LOG_LEVEL = os.getenv("LOG_LEVEL", "INFO").upper()
CACHE_DIR = os.getenv("CACHE_DIR", "./audio_cache")

logging.basicConfig(level=getattr(logging, LOG_LEVEL))
logger = logging.getLogger(__name__)

Path(CACHE_DIR).mkdir(parents=True, exist_ok=True)

SENT_SPLIT_RE = re.compile(r'(?<=[.!?,:])\s+')

# 内存缓存：单句缓存结构 {cache_key → { "audio": b64, "sr": 24000 }}
memory_cache = TTLCache(maxsize=MEMORY_CACHE_SIZE, ttl=18000)
cache_lock = threading.Lock()

executor = concurrent.futures.ThreadPoolExecutor(max_workers=16)

# 模型（常驻）
model_pipeline = None
model_lock = asyncio.Lock()

# 并发管理
request_semaphore = asyncio.Semaphore(MAX_CONCURRENT_REQUESTS)
current_requests: Dict[str, Dict] = {}


# ============================================================
# 模型加载
# ============================================================
async def load_model():
    global model_pipeline
    async with model_lock:
        if model_pipeline is None:
            try:
                model_pipeline = KPipeline(lang_code='a', device="cpu")
                logger.info("模型加载完成 (CPU)")
            except Exception as e:
                logger.error(f"模型加载失败: {e}")
                raise


@asynccontextmanager
async def lifespan(app: FastAPI):
    await load_model()
    yield
    logger.info("应用关闭中...")


app = FastAPI(lifespan=lifespan)


# CORS 允许所有来源
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)


# ============================================================
# 工具函数
# ============================================================

def split_sentences(text: str) -> List[str]:
    return [s.strip() for s in SENT_SPLIT_RE.split(text.strip()) if s.strip()]


def sentence_cache_key(sentence: str, voice: str, speed: float) -> str:
    raw = f"{sentence}|{voice}|{speed}"
    return hashlib.md5(raw.encode()).hexdigest()


def sentence_cache_path(key: str):
    return os.path.join(CACHE_DIR, f"{key}.wav")


def disk_meta_path(key: str):
    return os.path.join(CACHE_DIR, f"{key}.json")


def put_memory_cache(key: str, value: dict):
    with cache_lock:
        memory_cache[key] = value


def get_memory_cache(key: str) -> Optional[dict]:
    with cache_lock:
        return memory_cache.get(key)


async def load_sentence_from_disk(key: str) -> Optional[dict]:
    wav_path = sentence_cache_path(key)
    meta_path = disk_meta_path(key)

    if not Path(wav_path).exists() or not Path(meta_path).exists():
        return None

    async with aiofiles.open(meta_path, "r") as f:
        meta_text = await f.read()
        meta = json.loads(meta_text)

    audio, sr = await asyncio.get_event_loop().run_in_executor(
        executor, lambda: sf.read(wav_path)
    )
    with io.BytesIO() as buf:
        # sf.write(buf, audio, sr, "WAV")
        with sf.SoundFile(buf, 'w', sr, channels=audio.shape[1] if audio.ndim > 1 else 1, 
                  format='WAV', subtype='PCM_16') as f:
            f.write(audio)
        audio_b64 = base64.b64encode(buf.getvalue()).decode()

    meta["audio"] = audio_b64
    return meta


async def save_sentence_to_disk(key: str, audio: np.ndarray, sr: int, sentence: str):
    wav_path = sentence_cache_path(key)
    meta_path = disk_meta_path(key)

    await asyncio.get_event_loop().run_in_executor(
        executor, lambda: sf.write(wav_path, audio, sr, format="WAV")
    )

    meta = {
        "sentence": sentence,
        "sample_rate": sr,
    }

    async with aiofiles.open(meta_path, "w") as f:
        await f.write(json.dumps(meta))

    return meta


async def clean_disk_cache():
    files = sorted(Path(CACHE_DIR).glob("*.wav"), key=os.path.getmtime)
    if len(files) <= DISK_CACHE_SIZE:
        return
    remove_n = len(files) - DISK_CACHE_SIZE
    for f in files[:remove_n]:
        try:
            json_path = disk_meta_path(f.stem)
            f.unlink()
            if Path(json_path).exists():
                Path(json_path).unlink()
        except:
            pass


async def check_client_alive(client_id):
    while True:
        await asyncio.sleep(0.3)
        if client_id not in current_requests or current_requests[client_id].get("interrupt"):
            return False


# ============================================================
# 请求跟踪
# ============================================================

@app.middleware("http")
async def track_clients(request: Request, call_next):
    client_id = (
        request.query_params.get("client_id")
        or request.headers.get("X-Client-ID")
        or str(uuid.uuid4())
    )
    if request.url.path == "/generate":
        current_requests[client_id] = {"active": True}

    response = await call_next(request)

    if request.url.path == "/generate":
        current_requests.pop(client_id, None)

    return response


# ============================================================
# 主接口
# ============================================================

@app.post("/generate")
async def generate_audio_stream(data: Dict = Body(...)):
    async with request_semaphore:
        text = data.get("text", "")
        voice = data.get("voice", "af_heart")
        speed = float(data.get("speed", 1.0))
        client_id = data.get("client_id", str(uuid.uuid4()))

        if not text:
            raise HTTPException(400, "文本不能为空")

        await load_model()

        # 取消旧请求
        if client_id in current_requests:
            current_requests[client_id]["interrupt"] = True
            await asyncio.sleep(0.05)

        sentences = split_sentences(text)
        current_requests[client_id] = {"interrupt": False}

        async def stream():
            client_alive_task = asyncio.create_task(check_client_alive(client_id))

            try:
                for idx, sentence in enumerate(sentences):
                    if client_alive_task.done():
                        break

                    key = sentence_cache_key(sentence, voice, speed)

                    # 1) 内存缓存
                    cache_item = get_memory_cache(key)
                    if cache_item:
                        yield json.dumps({
                            "index": idx,
                            **cache_item
                        }).encode() + b"\n"
                        continue

                    # 2) 磁盘缓存
                    disk_item = await load_sentence_from_disk(key)
                    if disk_item:
                        put_memory_cache(key, disk_item)
                        yield json.dumps({
                            "index": idx,
                            **disk_item
                        }).encode() + b"\n"
                        continue

                    # 3) 无缓存 → 推理
                    generator = model_pipeline(sentence, voice=voice, speed=speed)

                    # 仅取第一个 chunk
                    for _, _, audio in generator:
                        sr = 24000
                        with io.BytesIO() as buf:
                            # sf.write(buf, audio, sr, "WAV")
                            with sf.SoundFile(buf, 'w', sr, channels=audio.shape[1] if audio.ndim > 1 else 1, 
                  format='WAV', subtype='PCM_16') as f:
                                f.write(audio)
                            audio_b64 = base64.b64encode(buf.getvalue()).decode()

                        item = {
                            "sentence": sentence,
                            "audio": audio_b64,
                            "sample_rate": sr
                        }

                        put_memory_cache(key, item)

                        # 保存磁盘缓存
                        await save_sentence_to_disk(key, audio, sr, sentence)
                        await clean_disk_cache()

                        yield json.dumps({
                            "index": idx,
                            **item
                        }).encode() + b"\n"
                        break

            finally:
                current_requests.pop(client_id, None)
                if not client_alive_task.done():
                    client_alive_task.cancel()

        return StreamingResponse(stream(), media_type="application/x-ndjson")


# ============================================================
# 其它接口
# ============================================================

@app.get("/clear-cache")
async def clear_cache():
    with cache_lock:
        memory_cache.clear()
    for f in Path(CACHE_DIR).glob("*"):
        f.unlink()
    return {"status": "success"}


@app.get("/cache-info")
async def get_cache_info():
    with cache_lock:
        mem_count = len(memory_cache)
    disk_files = list(Path(CACHE_DIR).glob("*.wav"))
    return {
        "memory_cache": mem_count,
        "disk_cache": len(disk_files)
    }


# ============================================================
# 启动
# ============================================================

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8028)