torch transforms inspired by sox effects

[vincentqb]

🚀 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 Merging plan from torchaudio-contrib #110 linking conversation: fast, general audio read keunwoochoi/torchaudio-contrib#31
• Some transforms touch on Merging items torchaudio-contrib that are not yet in torchaudio #259
• Performance comparison between libraries for audio loading here
• internal doc

[cyrta]

I can help. Which should I take as first ?

[vincentqb]

I can help. Which should I take as first ?

@cyrta -- The space is open right now :) are you interested in any in particular?

[engineerchuan]

We have implemented biquad, lowpass_biquad, highpass_biquad as part of the 8273c3f

[vincentqb]

VCTK depends on four transforms: gain, channels, rate, gain, dither. (Channels doesn't need its own transform, and its debatable for gain. We already have rate as Resample.)

From here:

        E.append_effect_to_chain("gain", ["-h"])
        E.append_effect_to_chain("channels", [1])
        E.append_effect_to_chain("rate", [16000])
        E.append_effect_to_chain("gain", ["-rh"])
        E.append_effect_to_chain("dither", ["-s"])

[vincentqb]

We have implemented biquad, lowpass_biquad, highpass_biquad as part of the 8273c3f

sox uses biquad for other transforms too, see bottom here.

[haideraltahan]

repeat can be done with a current PyTorch operation (Pad with replicate option) ?