Discrepancy betweenn_mels vs. mel bin size in MelSpectrogramLoss #102
Description
In the DAC paper, the authors discuss using varied mel bin sizes for the multi-scale spectrogram loss (pictured). In the implementation, n_mels is passed to the MelSpectrogramLoss function. The config file then takes n_mels, using the same numbers described in the paper (pictured below) as being the "mel bin sizes".
As far as I understand, the number of mel bins does not (necessarily) equal the mel bin width. Is this an error in the description in the paper, or in implementation? I also checked the Encodec paper which uses 64 as the number of bins, not the bin width. Thank you!
