Confusion about the deepfilternet2 paper

[uiuc-earable-computing]

Hi！First, really great work and thanks for open source and everything. But I have a few

1. For the loss in the paper, there is a spectrogram loss and a muli-resolution spectrogram loss. But I feel the multi-res loss could just include the spectrogram loss as one single window size, so why is it written separately in the paper? Did I miss something?
2. In the paper it says 60000HRTFs, I'm wondering if that's a typo? I feel it should be RIRs.

Thanks!

[Rikorose]

Hi,

1. The spectrogram loss is directly computed on the resulting spectrogram, while the multi-res loss is computed after an additional istft, and stft. Even though istft/stft are linear operations, this is not exactly the same (numerical differences, stft consistency problem).
2. Yes this should be RIRs and is corrected in an updated version which will be published in the IWAENC 2022 proceedings.

[github-actions[bot]]

This issue is stale because it has been open 90 days with no activity. Remove stale label or comment or this will be closed in 7 days.