nateanl
commented
1 year ago Hi @ahmed-fau, thanks for proposing the feature. Could you elaborate what kind of use cases for this adaptive filtering method, and what's the technical details behind the method you propose?
🚀 The feature
Hi there, I would to ask if there is any plan to support adaptive filtering functions to torchaudio library so that they can be used in native way for training/optimizing deep learning speech/audio models.
Currently, torchaudio supports 'lfilter' function similar to scipy.signal.lfilter. However, this function isn't easy to run with adaptive filters (i.e., filters whose transfer function changes over frames of an audio signals). Moreover, the lfilter function of torchaudio doesn't return zi (initial conditions for the filter delays), which disables simple refactoring to make it valid for adaptive filtering operations.
Motivation, pitch
Currently, lots of deep neural network models for speech/audio signal generation incorporate classical dsp adaptive filtering techniques such as LPC analysis/synthesis. It would be nice if torchaudio supports such functionality in order to make it ease to use for training these models with pytorch.
Alternatives
No response
Additional context
No response