Very slow running speed

#1
by yeyingxian - opened

I have a computer with i5-7400 and RTX 5060Ti. I installed voice-isolate in a venv enviroment. But the running speed is really very slow

image

Streaming mode used 87 seconds. Offline mode used 143 seconds. The source file sample_14.wav is only 10 seconds long

Thanks for trying it out, and for the timings.

Those numbers look like the model is running on your CPU rather than the RTX 5060 Ti. On Windows, a plain pip install pulls a CPU-only build of PyTorch, and the tool falls back to CPU when CUDA isn't available. For reference, the same file takes about a minute on a fast desktop CPU and about 3 seconds on a GPU.

Could you run this and paste the output?

voice-isolate-doctor --json

If the "device" line says cpu, reinstall PyTorch with CUDA support in the same venv:

pip install --force-reinstall torch torchaudio --index-url https://download.pytorch.org/whl/cu128

Then run voice-isolate-doctor --full again. It should report "inference runs on cuda", and sample_14.wav should finish much faster.

If it already says cuda and is still slow, please share the doctor output and we'll dig in. Smaller consumer GPUs aren't in our tested list yet, so your results are useful to us.

Thank you for your quickly reponse. I reinstalled torch with CUDA, then it runs much faster

image

Maybe you could add these points to the README, so that others can avoid making the same mistakes I did. I’m curious why streaming mode is slower than offline mode when running in CUDA?

Sign up or log in to comment