auraloss

A collection of audio-focused loss functions in PyTorch.

Setup

pip install auraloss

Usage

import torch
import auraloss

mrstft = auraloss.freq.MultiResolutionSTFTLoss()

input = torch.rand(8,1,44100)
target = torch.rand(8,1,44100)

loss = mrstft(input, target)

Loss functions

We categorize the loss functions as either time-domain or frequency-domain approaches. Additionally, we include perceptual transforms.

Loss function	Interface	Reference
Time domain
Error-to-signal ratio (ESR)	`auraloss.time.ESRLoss()`	Wright & Välimäki, 2019
DC error (DC)	`auraloss.time.DCLoss()`	Wright & Välimäki, 2019
Log hyperbolic cosine (Log-cosh)	`auraloss.time.LogCoshLoss()`	Chen et al., 2019
Signal-to-noise ratio (SNR)	`auraloss.time.SNRLoss()`
Scale-invariant signal-to-distortion ratio (SI-SDR)	`auraloss.time.SISDRLoss()`	Le Roux et al., 2018
Scale-dependent signal-to-distortion ratio (SD-SDR)	`auraloss.time.SDSDRLoss()`	Le Roux et al., 2018
Frequency domain
Spectral convergence	`auraloss.freq.SpectralConvergenceLoss()`	Arik et al., 2018
Log STFT magnitude	`auraloss.freq.LogSTFTMagnitudeLoss()`	Arik et al., 2018
Aggregate STFT	`auraloss.freq.STFTLoss()`	Arik et al., 2018
Aggregate Mel-scaled STFT	`auraloss.freq.MelSTFTLoss()`
Multi-resolution STFT	`auraloss.freq.MultiResolutionSTFTLoss()`	Yamamoto et al., 2019*
Random-resolution STFT	`auraloss.freq.RandomResolutionSTFTLoss()`	Steinmetz & Reiss, 2020
Sum and difference STFT loss	`auraloss.freq.SumAndDifferenceSTFTLoss()`	Steinmetz et al., 2020
Perceptual transforms
Sum and difference signal trasform	`auraloss.perceptual.SumAndDifference()`
FIR pre-emphasis filters	`auraloss.perceptual.FIRFilter()`	Wright & Välimäki, 2019

* Wang et al., 2019 also propose a multi-resolution spectral loss (that Engel et al., 2020 follow), but they do not include both the log magnitude (L1 distance) and spectral convergence terms, introduced in Arik et al., 2018, and then extended for the multi-resolution case in Yamamoto et al., 2019.

Examples

Currently we include an example using a set of the loss functions to train a TCN for modeling an analog dynamic range compressor. For details please refer to the details in examples/compressor. We provide pre-trained models, evaluation scripts to compute the metrics in the paper, as well as scripts to retrain models.

Development

We currently have no tests, but those will also be coming soon, so use caution at the moment. Future loss functions to be included will target neural network based perceptual losses, which tend to be a bit more sophisticated than those we have included so far.

If you are interested in adding a loss function please make a pull request.

Cite

If you use this code in your work please consider citing us.

@inproceedings{steinmetz2020auraloss,
    title={auraloss: {A}udio focused loss functions in {PyTorch}},
    author={Steinmetz, Christian J. and Reiss, Joshua D.},
    booktitle={Digital Music Research Network One-day Workshop (DMRN+15)},
    year={2020}}

donghaiyw/auraloss