SenseVoice-small-onnx\u662f\u4e00\u4e2a\u57fa\u4e8eONNX\u91cf\u5316\u7684\u591a\u8bed\u8a00\u8bed\u97f3\u8bc6\u522b\u6a21\u578b，\u4e13\u95e8\u9488\u5bf9\u8d44\u6e90\u53d7\u9650\u73af\u5883\u4f18\u5316\u3002\u8fd9\u4e2a\u6a21\u578b\u6700\u5927\u7684\u7279\u70b9\u662f\u5728\u4fdd\u6301\u9ad8\u7cbe\u5ea6\u7684\u540c\u65f6，\u5927\u5e45\u964d\u4f4e\u4e86\u8ba1\u7b97\u8d44\u6e90\u9700\u6c42，\u7279\u522b\u9002\u5408\u5728\u6ca1\u6709root\u6743\u9650\u7684\u670d\u52a1\u5668\u73af\u5883\u4e2d\u90e8\u7f72\u3002<\/p>\n
\u5728\u5b9e\u9645\u5e94\u7528\u4e2d，\u5f88\u591a\u5f00\u53d1\u8005\u548c\u4f01\u4e1a\u90fd\u4f1a\u9047\u5230\u8fd9\u6837\u7684\u56f0\u5883：\u9700\u8981\u90e8\u7f72AI\u670d\u52a1，\u4f46\u670d\u52a1\u5668\u6743\u9650\u53d7\u9650，\u65e0\u6cd5\u5b89\u88c5\u7cfb\u7edf\u7ea7\u4f9d\u8d56，\u6216\u8005\u786c\u4ef6\u8d44\u6e90\u6709\u9650\u3002SenseVoice-small-onnx\u6b63\u662f\u4e3a\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\u800c\u8bbe\u8ba1\u7684，\u5b83\u63d0\u4f9b\u4e86：<\/p>\n
\u5728\u53d7\u9650\u73af\u5883\u4e2d\u90e8\u7f72，\u6700\u5173\u952e\u7684\u662f\u786e\u4fdd\u6240\u6709\u4f9d\u8d56\u90fd\u80fd\u5728\u7528\u6237\u6743\u9650\u4e0b\u5b89\u88c5\u548c\u8fd0\u884c\u3002\u4ee5\u4e0b\u662f\u5b8c\u6574\u7684\u4f9d\u8d56\u5b89\u88c5\u6b65\u9aa4：<\/p>\n
# \u521b\u5efa\u865a\u62df\u73af\u5883（\u63a8\u8350）
\npython -m venv sensevoice_env
\nsource sensevoice_env\/bin\/activate<\/p>\n
# \u5b89\u88c5\u6838\u5fc3\u4f9d\u8d56
\npip install funasr-onnx gradio fastapi uvicorn soundfile jieba<\/p>\n
# \u9a8c\u8bc1\u5b89\u88c5
\npython -c "import funasr_onnx; print('funasr-onnx \u5b89\u88c5\u6210\u529f')"<\/p>\n
\u6ce8\u610f\u4e8b\u9879：<\/p>\n
\u5728\u65e0root\u6743\u9650\u73af\u5883\u4e2d，\u6a21\u578b\u6587\u4ef6\u9700\u8981\u5b58\u653e\u5728\u7528\u6237\u6709\u5199\u5165\u6743\u9650\u7684\u76ee\u5f55：<\/p>\n
import os
\nfrom pathlib import Path<\/p>\n
# \u5728\u7528\u6237\u76ee\u5f55\u4e0b\u521b\u5efa\u6a21\u578b\u5b58\u50a8\u8def\u5f84
\nhome_dir = Path.home()
\nmodel_dir = home_dir \/ "ai-models" \/ "sensevoice-small-onnx-quant"
\nmodel_dir.mkdir(parents=True, exist_ok=True)<\/p>\n
# \u8bbe\u7f6e\u6a21\u578b\u8def\u5f84\u73af\u5883\u53d8\u91cf
\nos.environ["MODEL_PATH"] = str(model_dir)<\/p>\n
\u5982\u679c\u670d\u52a1\u5668\u65e0\u6cd5\u76f4\u63a5\u4e0b\u8f7d\u6a21\u578b，\u53ef\u4ee5\u624b\u52a8\u4e0b\u8f7d\u540e\u4e0a\u4f20\u5230\u8be5\u76ee\u5f55：<\/p>\n
# \u624b\u52a8\u4e0b\u8f7d\u6a21\u578b\u6587\u4ef6（\u5982\u679c\u6709\u5916\u7f51\u6743\u9650）
\nwget -P ~\/ai-models\/sensevoice-small-onnx-quant\/ https:\/\/huggingface.co\/danieldong\/sensevoice-small-onnx-quant\/resolve\/main\/model_quant.onnx<\/p>\n
\u5728\u53d7\u9650\u73af\u5883\u4e2d，\u6211\u4eec\u9700\u8981\u4fee\u6539\u9ed8\u8ba4\u7684\u542f\u52a8\u811a\u672c，\u907f\u514d\u4f7f\u7528\u9700\u8981\u7279\u6743\u7aef\u53e3：<\/p>\n
# app_adapted.py – \u9002\u914d\u65e0root\u73af\u5883\u7684\u542f\u52a8\u811a\u672c
\nimport argparse
\nfrom funasr_onnx import SenseVoiceSmall
\nimport gradio as gr
\nimport os<\/p>\n
def create_app():
\n # \u4f7f\u7528\u7528\u6237\u76ee\u5f55\u4e0b\u7684\u6a21\u578b\u8def\u5f84
\n model_path = os.path.join(os.environ.get("MODEL_PATH", ".\/models"))<\/p>\n
model = SenseVoiceSmall(
\n model_path,
\n batch_size=5, # \u5728\u53d7\u9650\u73af\u5883\u4e2d\u51cf\u5c0fbatch size
\n quantize=True
\n )<\/p>\n
# \u521b\u5efaGradio\u754c\u9762
\n interface = gr.Interface(
\n fn=model.transcribe,
\n inputs=gr.Audio(type="filepath"),
\n outputs="text",
\n title="SenseVoice \u8bed\u97f3\u8bc6\u522b"
\n )<\/p>\n
return interface<\/p>\n
if __name__ == "__main__":
\n parser = argparse.ArgumentParser()
\n parser.add_argument("–host", default="0.0.0.0")
\n parser.add_argument("–port", default=8000, type=int) # \u4f7f\u7528\u975e\u7279\u6743\u7aef\u53e3
\n args = parser.parse_args()<\/p>\n
app = create_app()
\n app.launch(server_name=args.host, server_port=args.port)<\/p>\n
\u542f\u52a8\u670d\u52a1：<\/p>\n
# \u4f7f\u7528\u9002\u914d\u540e\u7684\u811a\u672c\u542f\u52a8
\npython app_adapted.py –host 0.0.0.0 –port 8000<\/p>\n
\u670d\u52a1\u542f\u52a8\u540e，\u53ef\u4ee5\u901a\u8fc7HTTP API\u8fdb\u884c\u8bed\u97f3\u8bc6\u522b：<\/p>\n
# api_client.py – \u5ba2\u6237\u7aef\u8c03\u7528\u793a\u4f8b
\nimport requests<\/p>\n
def transcribe_audio(audio_file_path, language="auto", use_itn=True):
\n """
\n \u8c03\u7528\u8bed\u97f3\u8bc6\u522bAPI<\/p>\n
Args:
\n audio_file_path: \u97f3\u9891\u6587\u4ef6\u8def\u5f84
\n language: \u8bed\u8a00\u4ee3\u7801，\u9ed8\u8ba4auto\u81ea\u52a8\u68c0\u6d4b
\n use_itn: \u662f\u5426\u4f7f\u7528\u9006\u6587\u672c\u6b63\u5219\u5316<\/p>\n
Returns:
\n \u8bc6\u522b\u7ed3\u679c\u6587\u672c
\n """
\n url = "http:\/\/localhost:8000\/api\/transcribe"<\/p>\n
with open(audio_file_path, 'rb') as f:
\n files = {'file': f}
\n data = {
\n 'language': language,
\n 'use_itn': str(use_itn).lower()
\n }<\/p>\n
response = requests.post(url, files=files, data=data)<\/p>\n
if response.status_code == 200:
\n return response.json()['text']
\n else:
\n raise Exception(f"\u8bc6\u522b\u5931\u8d25: {response.text}")<\/p>\n
# \u4f7f\u7528\u793a\u4f8b
\nresult = transcribe_audio("test_audio.wav", language="zh")
\nprint(f"\u8bc6\u522b\u7ed3\u679c: {result}")<\/p>\n
\u5728\u65e0\u7279\u6743\u73af\u5883\u4e2d，7860\u7aef\u53e3\u53ef\u80fd\u88ab\u5360\u7528，\u89e3\u51b3\u65b9\u6848：<\/p>\n
# \u67e5\u770b\u53ef\u7528\u7aef\u53e3
\nnetstat -tlnp 2>\/dev\/null | grep -v "0.0.0.0" | awk '{print $4}' | awk -F: '{print $NF}'<\/p>\n
# \u9009\u62e9\u5927\u4e8e1024\u7684\u672a\u4f7f\u7528\u7aef\u53e3
\npython app_adapted.py –port 8001<\/p>\n
\u5982\u679c\u670d\u52a1\u5668\u5185\u5b58\u6709\u9650，\u53ef\u4ee5\u8c03\u6574\u914d\u7f6e：<\/p>\n
# \u5185\u5b58\u4f18\u5316\u914d\u7f6e
\nmodel = SenseVoiceSmall(
\n model_path,
\n batch_size=2, # \u51cf\u5c0f\u6279\u5904\u7406\u5927\u5c0f
\n quantize=True,
\n device="cpu" # \u5f3a\u5236\u4f7f\u7528CPU，\u51cf\u5c11GPU\u5185\u5b58\u5360\u7528
\n)<\/p>\n
\u5982\u679c\u6a21\u578b\u4e0b\u8f7d\u5931\u8d25，\u53ef\u4ee5\u624b\u52a8\u5904\u7406：<\/p>\n
# 1. \u5728\u5176\u4ed6\u673a\u5668\u4e0b\u8f7d\u6a21\u578b
\n# 2. \u538b\u7f29\u540e\u4e0a\u4f20\u5230\u670d\u52a1\u5668
\n# 3. \u89e3\u538b\u5230\u7528\u6237\u76ee\u5f55
\ntar -xzf sensevoice-model.tar.gz -C ~\/ai-models\/<\/p>\n
\u5728\u53d7\u9650\u73af\u5883\u4e2d，\u6027\u80fd\u4f18\u5316\u5c24\u4e3a\u91cd\u8981：<\/p>\n
# performance_optimized.py – \u6027\u80fd\u4f18\u5316\u7248\u672c
\nimport time
\nfrom functools import lru_cache
\nfrom funasr_onnx import SenseVoiceSmall<\/p>\n
class OptimizedSenseVoice:
\n def __init__(self, model_path):
\n self.model = SenseVoiceSmall(
\n model_path,
\n batch_size=3,
\n quantize=True
\n )
\n self.last_audio_hash = None
\n self.last_result = None<\/p>\n
@lru_cache(maxsize=10)
\n def transcribe_cached(self, audio_path, language="auto"):
\n """\u5e26\u7f13\u5b58\u7684\u8f6c\u5f55\u529f\u80fd，\u907f\u514d\u91cd\u590d\u5904\u7406\u76f8\u540c\u97f3\u9891"""
\n return self.model([audio_path], language=language, use_itn=True)[0]<\/p>\n
def batch_transcribe(self, audio_paths, language="auto"):
\n """\u6279\u91cf\u5904\u7406，\u63d0\u9ad8\u6548\u7387"""
\n results = []
\n for path in audio_paths:
\n result = self.transcribe_cached(path, language)
\n results.append(result)
\n return results<\/p>\n
# \u4f7f\u7528\u4f18\u5316\u7248\u672c
\noptimized_model = OptimizedSenseVoice("~\/ai-models\/sensevoice-small-onnx-quant")<\/p>\n
\u901a\u8fc7\u672c\u6587\u7684\u9002\u914d\u65b9\u6848，\u5373\u4f7f\u5728\u65e0root\u6743\u9650\u7684\u53d7\u9650\u670d\u52a1\u5668\u73af\u5883\u4e2d，\u4e5f\u80fd\u6210\u529f\u90e8\u7f72\u548c\u8fd0\u884cSenseVoice-small-onnx\u8bed\u97f3\u8bc6\u522b\u670d\u52a1\u3002\u5173\u952e\u8981\u70b9\u5305\u62ec：<\/p>\n
\u4e0b\u4e00\u6b65\u5efa\u8bae：<\/p>\n
\u73b0\u5728\u4f60\u53ef\u4ee5\u5f00\u59cb\u5728\u5b9e\u9645\u7684\u53d7\u9650\u73af\u5883\u4e2d\u90e8\u7f72SenseVoice\u8bed\u97f3\u8bc6\u522b\u670d\u52a1，\u4eab\u53d7\u591a\u8bed\u8a00\u8bed\u97f3\u8f6c\u5199\u7684\u4fbf\u5229\u4e86\u3002<\/p>\n
\u83b7\u53d6\u66f4\u591aAI\u955c\u50cf<\/p>\n
\u60f3\u63a2\u7d22\u66f4\u591aAI\u955c\u50cf\u548c\u5e94\u7528\u573a\u666f？\u8bbf\u95ee CSDN\u661f\u56fe\u955c\u50cf\u5e7f\u573a，\u63d0\u4f9b\u4e30\u5bcc\u7684\u9884\u7f6e\u955c\u50cf，\u8986\u76d6\u5927\u6a21\u578b\u63a8\u7406\u3001\u56fe\u50cf\u751f\u6210\u3001\u89c6\u9891\u751f\u6210\u3001\u6a21\u578b\u5fae\u8c03\u7b49\u591a\u4e2a\u9886\u57df，\u652f\u6301\u4e00\u952e\u90e8\u7f72\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
SenseVoice-small-onnx\u5f00\u6e90\u8bed\u97f3\u8bc6\u522b\u6a21\u578b\u90e8\u7f72：\u65e0root\u6743\u9650\u670d\u52a1\u5668\u53d7\u9650\u73af\u5883\u9002\u914d
\n1. \u9879\u76ee\u6982\u8ff0\u4e0e\u6838\u5fc3\u4ef7\u503c
\nSenseVoice-small-onnx\u662f\u4e00\u4e2a\u57fa\u4e8eONNX\u91cf\u5316\u7684\u591a\u8bed\u8a00\u8bed\u97f3\u8bc6\u522b\u6a21\u578b，\u4e13\u95e8\u9488\u5bf9\u8d44\u6e90\u53d7\u9650\u73af\u5883\u4f18\u5316\u3002\u8fd9\u4e2a\u6a21\u578b\u6700\u5927\u7684\u7279\u70b9\u662f\u5728\u4fdd\u6301\u9ad8\u7cbe\u5ea6\u7684\u540c\u65f6，\u5927\u5e45\u964d\u4f4e\u4e86\u8ba1\u7b97\u8d44\u6e90\u9700\u6c42，\u7279\u522b\u9002\u5408\u5728\u6ca1\u6709root\u6743\u9650\u7684\u670d\u52a1\u5668\u73af\u5883\u4e2d\u90e8\u7f72\u3002
\n\u5728\u5b9e\u9645\u5e94\u7528\u4e2d，\u5f88\u591a\u5f00\u53d1\u8005\u548c\u4f01\u4e1a\u90fd\u4f1a\u9047\u5230\u8fd9\u6837\u7684\u56f0\u5883<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[7591,9288,1395],"topic":[],"class_list":["post-87126","post","type-post","status-publish","format-standard","hentry","category-server","tag-ai","tag-onnx","tag-1395"],"yoast_head":"\n