hailin
/
taixf
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix: run TTS synthesis in thread pool to avoid blocking event loop 

Also add size check for int8 model to skip LFS pointer files.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
hailin 2026-04-05 23:05:44 -07:00
parent 83cdf3396d
commit 9b2b875c2b
1 changed files with 10 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
backend/main/xiaozhi-server/core/providers/tts/sherpa_tts.py
View File

@ -1,5 +1,7 @@
import io import io
import os
import wave import wave
import asyncio
import numpy as np import numpy as np
import sherpa_onnx import sherpa_onnx
from config.logger import setup_logging from config.logger import setup_logging
@ -19,8 +21,7 @@ class TTSProvider(TTSProviderBase):
# 优先使用 int8 量化模型（更快） # 优先使用 int8 量化模型（更快）
model_file = f"{model_dir}/model.int8.onnx" model_file = f"{model_dir}/model.int8.onnx"
import os if not os.path.exists(model_file) or os.path.getsize(model_file) < 1024:
if not os.path.exists(model_file):
model_file = f"{model_dir}/model.onnx" model_file = f"{model_dir}/model.onnx"
num_threads = int(config.get("num_threads", 8)) num_threads = int(config.get("num_threads", 8))
@ -44,7 +45,8 @@ class TTSProvider(TTSProviderBase):
f"SherpaOnnxTTS 初始化完成: model_dir={model_dir}, sample_rate={self.sample_rate}, sid={self.sid}" f"SherpaOnnxTTS 初始化完成: model_dir={model_dir}, sample_rate={self.sample_rate}, sid={self.sid}"
) )
async def text_to_speak(self, text, output_file): def _generate_wav(self, text):
"""同步合成，在线程池中调用"""
audio = self.tts.generate(text, sid=self.sid, speed=self.speed) audio = self.tts.generate(text, sid=self.sid, speed=self.speed)
samples = np.array(audio.samples, dtype=np.float32) samples = np.array(audio.samples, dtype=np.float32)
pcm = (samples * 32767).astype(np.int16) pcm = (samples * 32767).astype(np.int16)
@ -55,7 +57,11 @@ class TTSProvider(TTSProviderBase):
wf.setsampwidth(2) wf.setsampwidth(2)
wf.setframerate(self.sample_rate) wf.setframerate(self.sample_rate)
wf.writeframes(pcm.tobytes()) wf.writeframes(pcm.tobytes())
wav_data = wav_io.getvalue() return wav_io.getvalue()
async def text_to_speak(self, text, output_file):
loop = asyncio.get_event_loop()
wav_data = await loop.run_in_executor(None, self._generate_wav, text)
if output_file: if output_file:
with open(output_file, "wb") as f: with open(output_file, "wb") as f: