auraloss

A collection of audio-focused loss functions in PyTorch.

[PDF]

Setup

pip install auraloss

If you want to use MelSTFTLoss() or FIRFilter() you will need to specify the extra install (librosa and scipy).

pip install auraloss[all]

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)

NEW: Perceptual weighting with mel scaled spectrograms.

bs = 8
chs = 1
seq_len = 131072
sample_rate = 44100

# some audio you want to compare
target = torch.rand(bs, chs, seq_len)
pred = torch.rand(bs, chs, seq_len)

# define the loss function
loss_fn = auraloss.freq.MultiResolutionSTFTLoss(
    fft_sizes=[1024, 2048, 8192],
    hop_sizes=[256, 512, 2048],
    win_lengths=[1024, 2048, 8192],
    scale="mel",
    n_bins=128,
    sample_rate=sample_rate,
    perceptual_weighting=True,
)

# compute
loss = loss_fn(pred, target)

Citation

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}
}

Loss functions

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

* 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.

There are some more advanced things you can do based upon the STFTLoss class. For example, you can compute both linear and log scaled STFT errors as in Engel et al., 2020. In this case we do not include the spectral convergence term.

stft_loss = auraloss.freq.STFTLoss(
    w_log_mag=1.0, 
    w_lin_mag=1.0, 
    w_sc=0.0,
)

There is also a Mel-scaled STFT loss, which has some special requirements. This loss requires you set the sample rate as well as specify the correct device.

sample_rate = 44100
melstft_loss = auraloss.freq.MelSTFTLoss(sample_rate, device="cuda")

You can also build a multi-resolution Mel-scaled STFT loss with 64 bins easily. Make sure you pass the correct device where the tensors you are comparing will be.

loss_fn = auraloss.freq.MultiResolutionSTFTLoss(
    scale="mel", 
    n_bins=64,
    sample_rate=sample_rate,
    device="cuda"
)

If you are computing a loss on stereo audio you may want to consider the sum and difference (mid/side) loss. Below we have shown an example of using this loss function with the perceptual weighting and mel scaling for further perceptual relevance.

target = torch.rand(8, 2, 44100)
pred = torch.rand(8, 2, 44100)

loss_fn = auraloss.freq.SumAndDifferenceSTFTLoss(
    fft_sizes=[1024, 2048, 8192],
    hop_sizes=[256, 512, 2048],
    win_lengths=[1024, 2048, 8192],
    perceptual_weighting=True,
    sample_rate=44100,
    scale="mel",
    n_bins=128,
)

loss = loss_fn(pred, target)

Development

Run tests locally with pytest.

python -m pytest