torch transforms inspired by sox effects

🚀 Feature

Convert sox effects into transforms. Ideally, we provide enough building blocks to make those transforms easy to assemble.

• allpass (with biquad)
• band (with biquad)
• bandpass/bandreject (with biquad)
• bass (with biquad)
• treble (with biquad)
• bend
• biquad
• ~channels (we already decided downmix, etc, is easy to do by hand)~
• chorus
• compand/mcompand
• contrast
• dcshift
• deemph (with biquad, related to this)
• delay
• dither
• downsample/upsample (available with resample)
• earwax
• echo/echos
• equalizer (with biquad, see also here)
• fade
• fir
• flanger
• gain
• highpass/lowpass (with biquad)
• hilbert
• ladspa
• loudness
• mcompand
• noiseprof/noisered
• ~norm (duplicate of gain)~
• oops (simply the difference between two channels)
• overdrive
• ~pad/trim (easy to do without transforms)~
• phaser
• pitch
• ~rate (resample already available)~
• ~remix (we already decided downmix, etc, is easy to do by hand)~
• ~repeat (tensor.repeat or pad with replicate)~
• reverb
• ~reverse~
• riaa (with biquad)
• silence
• sinc
• spectrogram
• ~speed (change sample rate)~
• stretch/tempo (vocoder functional available)
• splice
• stat/stats (only prints information about the waveform)
• ~swap~
• synth
• tremolo
• vad (in example)
• vol

Crossed out: can be done in a simple way with current pytorch operations, so we don’t need a dedicated transform. Feel free to point out any updates recommended to this list.

Additional context

• Remove dependency on sox in VCTK
• Open files in format such as mp3, etc (see here and here)
• Comment from #110 linking keunwoochoi/torchaudio-contrib#31
• Some transforms touch on #259
• Performance comparison between libraries for audio loading here
• internal doc

[mthrok]

FYI: Regarding the speed, I did simple profiling with the following script and it tells that the most time consuming part is the iteration in lfilter

Script

import pprofile

import torch
import torchaudio
import torchaudio.functional as F


def profile_biquad():
    waveform, sr = torchaudio.load('test/assets/100Hz_44100Hz_16bit_05sec.wav', normalization=True)

    prof = pprofile.Profile()
    with prof():
        F.equalizer_biquad(waveform, sr, 3000, 1, 0.707)
    prof.print_stats()


profile_biquad()

Most time consuming parts

   720|    220501|       4.4464|   2.0165e-05| 18.97%|    for i_sample, o0 in enumerate(input_signal_windows.t()):
(call)|         1|  7.86781e-05|  7.86781e-05|  0.00%|# /scratch/moto/pytorch/torch/tensor.py:460 __iter__
(call)|    220500|      2.72458|  1.23564e-05| 11.62%|# /scratch/moto/pytorch/torch/tensor.py:474 <lambda>
   721|    220500|      2.80982|   1.2743e-05| 11.99%|        windowed_output_signal = padded_output_waveform[:, i_sample:(i_sample + n_order)]
   722|    220500|      4.92106|  2.23177e-05| 21.00%|        o0.sub_(torch.mv(windowed_output_signal, a_coeffs_flipped))
   723|    220500|      3.72974|  1.69149e-05| 15.91%|        o0.div_(a_coeffs[0])
   724|         0|            0|            0|  0.00%|
   725|    220500|      4.77714|   2.1665e-05| 20.38%|        padded_output_waveform[:, i_sample + n_order - 1] = o0

[tomassosorio]

Hey! I will try to remove the pitch dependency of sox, if is no problem @vincentqb 😃