auraloss

library0.4.0pypypi✓ verified 85d ago

Auraloss is a collection of audio-focused loss functions implemented in PyTorch, designed for tasks like audio synthesis, source separation, and speech enhancement. It provides specialized losses such as Mel-spectrogram, multi-resolution STFT, and perceptual losses. The current stable version is 0.4.0, and new features and improvements are added periodically, with releases typically following significant development milestones.

ai-ml data

pip install auraloss

github ↗package ↗

INSTALL

IMPORT

SIG · AURALOSS

auraloss

ai-mlpythonv0.4.0

Install

—

Import

—

Disk

—

Pass rate

0/ 10

Env Coverage0 / 10

glibc

3.9–3.13

musl

3.9–3.13

Install & Compatibility

Where this runs

tested against v0.4.0 · pip install

no network on importno background threads

Install × environment matrix

Each cell = how many times install + import succeeded across repeated harness runs. Partial = flaky.
glibc = Debian/Ubuntu slim · musl = Alpine Linux

musl

glibc

py 3.10

✕ build_error

2/4 runs

py 3.11

✕ build_error

2/4 runs

py 3.12

✕ build_error

2/4 runs

py 3.13

✕ build_error

2/4 runs

py 3.9

✕ build_error

✕ timeout

Code

Verified usage

Verified import paths — ran on the pinned version, not inferred.

MultiResolutionSTFTLoss

✓ from auraloss.freq import MultiResolutionSTFTLoss

MelSpectrogramLoss

✓ from auraloss.freq import MelSpectrogramLoss

PerceptualLoss

✓ from auraloss.perceptual import PerceptualLoss

SpectralConvergenceLoss

✓ from auraloss.freq import SpectralConvergenceLoss

This quickstart demonstrates how to instantiate and use the MultiResolutionSTFTLoss, a common and powerful loss function in auraloss. It generates dummy audio tensors and calculates the loss between them, showcasing the basic API for most loss functions in the library. Ensure PyTorch is installed and CUDA is available for GPU acceleration.

import torch from auraloss.freq import MultiResolutionSTFTLoss device = 'cuda' if torch.cuda.is_available() else 'cpu' # Dummy input/target tensors (e.g., 10 seconds of mono audio at 16kHz) # Batch size B, Channels C, Samples S input_audio = torch.randn(2, 1, 160000, device=device) target_audio = torch.randn(2, 1, 160000, device=device) # Initialize Multi-Resolution STFT Loss # The paper recommends a set of default parameters for MR-STFT Loss # consisting of 3 STFT magnitudes, with varying window sizes and hop sizes. # auraloss.freq.MultiResolutionSTFTLoss provides these defaults. mr_stft_loss = MultiResolutionSTFTLoss().to(device) # Compute the loss loss = mr_stft_loss(input_audio, target_audio) print(f"Computed MR-STFT Loss: {loss.item()}")

Debug

Known issues

gotchaAll auraloss functions expect input and target tensors to be 3-dimensional (Batch, Channels, Samples). A common mistake is to pass 2D (Batch, Samples) or 1D (Samples) tensors.

fix

Reshape your tensors to `(B, C, S)` using `tensor.unsqueeze(1)` for mono audio or `tensor.view(B, C, S)` for multi-channel audio where `C` is the channel dimension.

affects: >=0.1.0

gotchaEnsure input and target tensors are on the same device (CPU/GPU) as the loss function instance. Mismatched devices will lead to `RuntimeError: Expected all tensors to be on the same device`.

fix

Move all tensors and the loss module to the target device: `input_audio = input_audio.to(device)`, `target_audio = target_audio.to(device)`, `loss_fn = loss_fn.to(device)`.

affects: >=0.1.0

gotchaThe STFT-based losses (e.g., MultiResolutionSTFTLoss, MelSpectrogramLoss) rely on `torchaudio`'s STFT implementation, which might have specific requirements for tensor dtypes (typically `torch.float32` or `torch.float64`). Using other dtypes like `torch.float16` might cause issues or unexpected behavior.

fix

Ensure your audio tensors are `torch.float32` or `torch.float64`. You can cast them using `tensor.to(dtype=torch.float32)`.

affects: >=0.1.0

breakingPrior to v0.3.0, some loss functions like `STFTLoss` and `MelSTFTLoss` were directly in `auraloss.loss.STFTLoss` or `auraloss.loss.MelSTFTLoss`. They were later refactored into `auraloss.freq` and renamed.

fix

Update your imports: `from auraloss.freq import MultiResolutionSTFTLoss` for the multi-res version, or `from auraloss.freq import MelSpectrogramLoss`. The original `STFTLoss` was superseded by `MultiResolutionSTFTLoss` for improved performance and robustness.

affects: <0.3.0

Errors

Common errors & fixes

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0!

Input audio tensors or target audio tensors are on a different device (CPU/GPU) than the initialized auraloss module.

fix

Ensure all tensors and the loss module are moved to the same device: `loss_fn = loss_fn.to(device)`, `input_audio = input_audio.to(device)`, `target_audio = target_audio.to(device)`.

ValueError: Expected input to be a 3D tensor, got 2D tensor

auraloss functions expect input audio to be in (Batch, Channels, Samples) format, but received a 2D tensor (e.g., Batch, Samples).

fix

For mono audio, add a channel dimension using `tensor.unsqueeze(1)`: `input_audio = input_audio.unsqueeze(1)`.

TypeError: 'module' object is not callable

Attempting to call the auraloss module directly (e.g., `auraloss.freq(input, target)`) instead of an instantiated loss class.

fix

You need to import and instantiate a specific loss class first, then call its instance: `from auraloss.freq import MultiResolutionSTFTLoss; mr_loss = MultiResolutionSTFTLoss(); loss = mr_loss(input, target)`.

Upgrade

Version history

0.4.0latest on PyPI · released Apr 21, 2023

Audit

Dependencies

torchrequiredCore deep learning framework

torchaudiorequiredAudio processing utilities, particularly for STFT/Mel-spectrogram computations

Agent activity

40 hits · last 30 days

node

OpenAI (training)

Resources

githubhttps://github.com/csteinmetz1/auraloss ↗

packagehttps://pypi.org/project/auraloss/ ↗