sp-nitech / diffsptk Goto Github PK

It would be useful to be able to tweak make_filter_banks parameters in PQMF and IPQMF creation. In particular, these are missing:

    n_iter=100,
    step_size=1e-2,
    decay=0.5,
    eps=1e-6,

As the title said, when using MLSA with mode="freq-domain", fft_length=1024, alpha=0, it triggers the assertion error at:

This is due to the argument n_fft of MelGeneralizedCepstrumToSpectrum is not exposed to FrequencyDomainFIRFilter so it's always 512.

Is it intentional?
I propose to add n_fft as an argument to FrequencyDomainFIRFilter.

BTW, the repository description on GitHub is "A differential version of SPTK".
I guess it should be "A differentiable ..." which you also used in the README file.

Hi, as the title said, is there any plan to provide a functional module, like PyTorch did, for some state-less transformations?
For example, conversion functions like lar2par, par2lpc, levdur, and so on depend on an order parameter that can be inferred from the input data.
Having them as pure functions would be great without creating a nn.Module instance.
It would also be cool to have equivalent functional counterparts of other transforms, with default arguments that work out of the box when chaining the functions, similar to how librosa does.