Gammatone Filterbanks and Gammatone Spectrogram

🚀 The feature

The development of a Gammatone Filterbank and Gammatone Spectrogram in torchaudio to provide another method of audio feature extraction

The Gammatone Filterbank uses Equal Rectangular Bandwith scale (ERB) to scale the frequency range which are closely associated with the human auditory filters found within the cochlea. Furthermore, a gammatone filter is used which is a non-uniform band-pass filter which is suggested to be an improved method to filters such as the roex (rounded-exponential filter) used in ISO 532-2. Ultimately, this feature aims to be an improvement on the Mel Scale and filterbank which uses triangular filtering.

The gammatone filter will use Scipy to obtain the required filters and associated frequency response. The user will be able to obtain both the Gammatone Filterbanks and the Gammatone Spectrogram through a single transformation on both mono and stereo sources (or any number of channels).

Motivation, pitch

The motivation behind this development is to provide an audio feature extraction which is more closely associated with the understanding of the human auditory filters found within the cochlea. This will use the ERB Scale and Gammatone filters to create a filterbank and spectrogram. The Gammatone filterbank and Gammatone Spectrogram will be a useful feature for audio classification purposes.

Alternatives

The only direct alternative is found in Matlab which implements the Gammatone Filterbank.

Additional context

The features have already been tested on a real-world environemntal sound classification issue with improved results. Furthermore, the Gammatone Spectrogram produces a cleaner spectrogram which has been attached. Note that the inputs for each spectrogram function (i.e. n_fft, win_len, hop_len, n_bands) are the same. The audio sample is that of a person speaking at a distance from the microphone.

No files, tests, or entry points are named. Start by reviewing torchaudio's existing filterbank and spectrogram transformations, then determine the API and dependency approach for ERB-scaled Gammatone filterbanks and spectrograms. Done should include documented behavior and coverage for mono, stereo, and multi-channel inputs.