Proper STFT configuration for DCUNet

asteroid-team / asteroid

The PyTorch-based audio source separation toolkit for researchers

https://asteroid-team.github.io/

MIT License

2.27k stars 423 forks source link

Proper STFT configuration for DCUNet #400

Closed jonashaag closed 3 years ago

jonashaag commented 3 years ago

🐛 Bug

For DCUNet, we currently use a 512/256 STFT. In the paper they use 1024/256. My own experiments show a large drop of performance for 512 FFT, and a large drop for hop >= 1/2 FFT. I don't see a performance increase with 2048/512.

I suggest to change the default to 1024/256.

Note that my tests were done with torch_stft_fb.

mpariente commented 3 years ago

Yes, let's use the defaults as in the paper, sure.

mpariente commented 3 years ago

Fixed by #406