241 lines
7.9 KiB
Python
241 lines
7.9 KiB
Python
#!/usr/bin/env python3
|
|
"""
|
|
Genie TTS 极速服务 - 支持文本分块并行合成 + 线程池预热
|
|
适配 ARM 大小核架构（Cortex-X925 + Cortex-A725, 20核），推荐 MAX_WORKERS=8
|
|
"""
|
|
|
|
import os
|
|
os.environ["GENIE_DATA_DIR"] = r"/home/admin/minor_agent/modelscope/hub/models/genie_tts/GenieData"
|
|
|
|
import io
|
|
import tempfile
|
|
import re
|
|
import wave
|
|
import multiprocessing
|
|
from concurrent.futures import ThreadPoolExecutor, as_completed
|
|
from contextlib import asynccontextmanager
|
|
|
|
from fastapi import FastAPI, Query
|
|
from fastapi.responses import Response
|
|
import genie_tts as genie
|
|
|
|
# ======================== 配置区域 ========================
|
|
DEFAULT_CHARACTER = "minor"
|
|
DEFAULT_ONNX_MODEL_DIR = r"/home/admin/minor_agent/modelscope/models/genie_tts/CharacterModels/v2ProPlus/feibi/tts_models"
|
|
DEFAULT_REFERENCE_AUDIO = "/home/admin/minor_agent/modelscope/models/genie_tts/CharacterModels/v2ProPlus/minor/minor.wav"
|
|
DEFAULT_REFERENCE_TEXT = "你好你好你好，这是一段...测试录音~"
|
|
DEFAULT_LANGUAGE = "zh"
|
|
TMP_DIR = "/dev/shm" # 内存文件系统，减少磁盘 I/O
|
|
|
|
# 分块与并行参数
|
|
MIN_EFFECTIVE_CHARS = 5 # 每个块至少包含的有效字符数
|
|
MAX_WORKERS = 4 # 建议 4~8，ARM 大小核不要超过大核数（10）
|
|
POOL_WARMUP = True # 是否在线程池创建后立即预热
|
|
# =========================================================
|
|
|
|
# ---------- 全局线程池（应用生命周期内复用） ----------
|
|
tts_pool: ThreadPoolExecutor = None
|
|
|
|
# ---------- 正则工具 ----------
|
|
SENT_SPLIT_PATTERN = re.compile(r'(?<=[。！？；，、])')
|
|
REMOVE_PUNCT_PATTERN = re.compile(r'[。！？；，、：；“”（）《》【】\-\_\—\~\.\,\!\?\s·]')
|
|
|
|
def effective_length(text: str) -> int:
|
|
"""计算有效字符数（去除标点、连接符、空格）"""
|
|
return len(REMOVE_PUNCT_PATTERN.sub('', text))
|
|
|
|
def split_text(text: str) -> list[str]:
|
|
"""将文本切分为多个语义块"""
|
|
raw_chunks = [c.strip() for c in SENT_SPLIT_PATTERN.split(text) if c.strip()]
|
|
if not raw_chunks:
|
|
return [text]
|
|
|
|
merged = []
|
|
for chunk in raw_chunks:
|
|
if merged and effective_length(chunk) < MIN_EFFECTIVE_CHARS:
|
|
merged[-1] += chunk
|
|
else:
|
|
merged.append(chunk)
|
|
|
|
if len(merged) == 1:
|
|
return merged
|
|
|
|
if effective_length(merged[-1]) < MIN_EFFECTIVE_CHARS and len(merged) > 1:
|
|
merged[-2] += merged[-1]
|
|
merged.pop()
|
|
|
|
return merged
|
|
|
|
# ---------- 音频拼接 ----------
|
|
def merge_wavs(file_paths: list[str]) -> bytes:
|
|
"""按路径顺序拼接多个 WAV 文件"""
|
|
if len(file_paths) == 1:
|
|
with open(file_paths[0], 'rb') as f:
|
|
return f.read()
|
|
|
|
data_parts = []
|
|
params = None
|
|
for path in file_paths:
|
|
with wave.open(path, 'rb') as wf:
|
|
if params is None:
|
|
params = wf.getparams()
|
|
data_parts.append(wf.readframes(wf.getnframes()))
|
|
|
|
buf = io.BytesIO()
|
|
with wave.open(buf, 'wb') as wf_out:
|
|
wf_out.setparams(params)
|
|
for frames in data_parts:
|
|
wf_out.writeframes(frames)
|
|
return buf.getvalue()
|
|
|
|
# ---------- 预热函数 ----------
|
|
def _warmup_task(idx: int):
|
|
"""空任务用于预热线程"""
|
|
pass
|
|
|
|
def warmup_pool(pool: ThreadPoolExecutor, num_workers: int):
|
|
"""向线程池提交空任务以预热所有线程"""
|
|
futures = []
|
|
for i in range(num_workers):
|
|
futures.append(pool.submit(_warmup_task, i))
|
|
# 等待所有预热任务完成
|
|
for f in futures:
|
|
f.result()
|
|
print(f"✅ 线程池预热完成 (workers={num_workers})")
|
|
|
|
# ---------- 并行 TTS（使用全局线程池） ----------
|
|
def parallel_tts(chunks: list[str]) -> list[str]:
|
|
"""
|
|
多线程并行生成所有语音块，返回按原始顺序的临时文件路径列表。
|
|
"""
|
|
tmp_paths = []
|
|
for _ in chunks:
|
|
tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False, dir=TMP_DIR)
|
|
tmp.close()
|
|
tmp_paths.append(tmp.name)
|
|
|
|
try:
|
|
# 使用全局预热的线程池
|
|
future_to_idx = {}
|
|
for idx, chunk in enumerate(chunks):
|
|
future = tts_pool.submit(
|
|
genie.tts,
|
|
character_name=DEFAULT_CHARACTER,
|
|
text=chunk,
|
|
play=False,
|
|
save_path=tmp_paths[idx]
|
|
)
|
|
future_to_idx[future] = idx
|
|
|
|
for future in as_completed(future_to_idx):
|
|
future.result() # 传递异常
|
|
except:
|
|
for p in tmp_paths:
|
|
if os.path.exists(p):
|
|
os.unlink(p)
|
|
raise
|
|
return tmp_paths
|
|
|
|
# ---------- 应用生命周期 ----------
|
|
@asynccontextmanager
|
|
async def lifespan(app: FastAPI):
|
|
global tts_pool
|
|
|
|
print("🚀 正在启动 Genie TTS 极简服务...")
|
|
try:
|
|
# 1. 加载模型
|
|
genie.load_character(
|
|
character_name=DEFAULT_CHARACTER,
|
|
onnx_model_dir=DEFAULT_ONNX_MODEL_DIR,
|
|
language=DEFAULT_LANGUAGE
|
|
)
|
|
genie.set_reference_audio(
|
|
character_name=DEFAULT_CHARACTER,
|
|
audio_path=DEFAULT_REFERENCE_AUDIO,
|
|
audio_text=DEFAULT_REFERENCE_TEXT,
|
|
language=DEFAULT_LANGUAGE
|
|
)
|
|
|
|
# 2. 创建并预热全局线程池
|
|
print(f"🔥 正在创建线程池并预热 (max_workers={MAX_WORKERS})...")
|
|
tts_pool = ThreadPoolExecutor(max_workers=MAX_WORKERS)
|
|
if POOL_WARMUP:
|
|
warmup_pool(tts_pool, MAX_WORKERS)
|
|
|
|
print("✅ Genie TTS 服务启动成功！")
|
|
print(f" - 线程池: max_workers={MAX_WORKERS}, 已预热")
|
|
print("👉 调用方式：http://localhost:8000/tts?text=你要说的话")
|
|
except Exception as e:
|
|
print(f"❌ 服务启动失败: {e}")
|
|
raise e
|
|
|
|
yield
|
|
|
|
print("\n🛑 正在关闭服务...")
|
|
# 清理线程池
|
|
if tts_pool:
|
|
print(" 关闭线程池...")
|
|
tts_pool.shutdown(wait=True)
|
|
tts_pool = None
|
|
|
|
# 卸载模型
|
|
try:
|
|
genie.unload_character(DEFAULT_CHARACTER)
|
|
print("✅ 资源已释放")
|
|
except:
|
|
pass
|
|
|
|
app = FastAPI(lifespan=lifespan)
|
|
|
|
# ---------- 极简 API ----------
|
|
@app.get("/tts")
|
|
async def text_to_speech(text: str = Query(..., description="待合成的文本")):
|
|
if not text.strip():
|
|
return Response(content="文本不能为空", status_code=400, media_type="text/plain")
|
|
|
|
chunks = split_text(text)
|
|
|
|
try:
|
|
if len(chunks) == 1:
|
|
# 单块直接合成，避免多线程开销
|
|
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False, dir=TMP_DIR) as tmp:
|
|
tmp_path = tmp.name
|
|
try:
|
|
genie.tts(
|
|
character_name=DEFAULT_CHARACTER,
|
|
text=chunks[0],
|
|
play=False,
|
|
save_path=tmp_path
|
|
)
|
|
with open(tmp_path, "rb") as f:
|
|
audio_bytes = f.read()
|
|
return Response(content=audio_bytes, media_type="audio/wav")
|
|
finally:
|
|
if os.path.exists(tmp_path):
|
|
os.unlink(tmp_path)
|
|
else:
|
|
# 多块并行合成（使用预热好的线程池）
|
|
tmp_paths = parallel_tts(chunks)
|
|
try:
|
|
merged_audio = merge_wavs(tmp_paths)
|
|
return Response(content=merged_audio, media_type="audio/wav")
|
|
finally:
|
|
for p in tmp_paths:
|
|
if os.path.exists(p):
|
|
os.unlink(p)
|
|
except Exception as e:
|
|
return Response(content=f"生成失败: {str(e)}", status_code=500, media_type="text/plain")
|
|
|
|
@app.get("/health")
|
|
async def health():
|
|
return {
|
|
"status": "ok",
|
|
"pool_active": tts_pool is not None,
|
|
"max_workers": MAX_WORKERS
|
|
}
|
|
|
|
# ---------- 启动入口 ----------
|
|
if __name__ == "__main__":
|
|
import uvicorn
|
|
uvicorn.run(app, host="0.0.0.0", port=8000)
|