bfloat16
bfloat16
推荐用faster_whisper，结合VAD最大程度防止whisper抽风重复（ whisper.cpp貌似也行，但集成不方便 
> The only problem is as stated in issue #808, it is not recommended to use MPS for training on MacOS. Already modified
切片太长，共享显存占用严重
1.Try to download FFmpeg from https://www.gyan.dev/ffmpeg/builds/ ,and configure environment variables 2.run pip install ffmpeg-python Hope it helps you :)
打标不干净喵~
After stripping stft and istft, both onnx and trt engine can be exported successfully. onnx is not tested, and trt engine has serious performance problems (seems to be caused by...
> > After stripping stft and istft, both onnx and trt engine can be exported successfully. onnx is not tested, and trt engine has serious performance problems (seems to be...
torch要安装CPU版本吧，包里面的应该是cuda版本的（
逆了一下，从链上拿代码执行，这不此地无银三百两
用这个工具https://github.com/lucasg/Dependencies 检查D:\GPT-SoVITS\runtime\lib\site-packages\torch\lib\fbgemm.dll缺失了哪个dll，然后手动下载，放到fbgemm.dll旁边