Regarding the problem that the audio file length cannot be compressed

[rezzsl]

Hello, I would like to ask, when I am compressing some audio files, some of the audio files are short, and the code requires that the time length must be a multiple of the overlap, so it cannot be processed. How can I deal with this situation to successfully compress it? Thank you.

[yzGuu830]

Hi @rezzsl, thank you for bringing this up. Our implementation actually requires that $T$ be multiple of $(\text{overlap} \cdot \text{patchsize})$ (for patchify and frame overlap), and that $L$ be multiple of $\text{hoplen}$ (to ensure lossless inverse STFT).

For practical use of our pretrained models, I suggest either 1) truncating (with at most 20ms loss), or 2) padding (then truncating back to original length) the input audio length $L$ to be in the form $80 (4k-1), k \in \mathbb{Z}_+$. In streaming setup, you can buffer incoming audio until it satisfies the form. Alternatively, you could modify PatchEmbed and PathDeEmbed modules to include conditional padding patterns that can handle any input length.

[lixinghe1999]

need add one line to compensate
x = torch.nn.functional.pad(x, (0, 320 - x.shape[0] % 320 - 80), mode='constant', value=0)