TorchLibrosa: PyTorch Implementation of Librosa

0.1.0 · active · verified Thu Apr 16

TorchLibrosa provides a PyTorch implementation of core `librosa` audio feature extraction functions, enabling GPU acceleration for tasks such as spectrogram and mel-spectrogram computation. This is particularly beneficial for deep learning pipelines that require faster feature generation on GPUs during training and evaluation. The library aims for numerical results almost identical to CPU-based `librosa` (difference less than 1e-5). The current version is 0.1.0, with an infrequent release cadence; the latest release was in February 2023.

Common errors

```
ModuleNotFoundError: No module named 'torch'
```
cause `torchlibrosa` implicitly depends on `torch` but does not list it in its `install_requires`.

fix Install PyTorch separately: `pip install torch` (or the appropriate version for your CUDA setup).
```
TypeError: pad_center() takes 1 positional argument but 2 were given
```
cause This error has been reported in older versions (e.g., 0.0.9) when initializing `STFT` or `Spectrogram`, indicating a potential mismatch in expected arguments for internal padding functions, possibly due to `librosa` or `torch` version incompatibilities.

fix Update `torchlibrosa` to version 0.1.0 and ensure your `librosa` dependency is `>=0.9.0` as specified. If the issue persists, try updating PyTorch to a recent stable version.
```
RuntimeError: mat1 and mat2 shapes cannot be multiplied
```
cause This is a common PyTorch error, often occurring when input tensors to `torchlibrosa` modules (which internally use matrix multiplications) do not have the expected dimensions or shape.

fix Carefully review the expected input shape for the `torchlibrosa` module you are using (e.g., `Spectrogram` expects `(batch_size, samples)`). Ensure your input tensor matches these requirements, potentially using `unsqueeze` or `permute` to reshape.

Warnings

gotcha PyTorch is a fundamental dependency for `torchlibrosa`, but it is not explicitly listed in the `install_requires` of the `setup.py`. This can lead to a `ModuleNotFoundError` if PyTorch is not installed separately.
Fix: Ensure PyTorch is installed manually via `pip install torch` (and `torchvision`, `torchaudio` if needed) or by specifying it in your project's `requirements.txt`.
gotcha While `torchlibrosa` aims to provide 'almost identical features' to `librosa`, a 'numerical difference less than 1e-5' is explicitly stated. Users expecting bit-for-bit identical results to `librosa` on CPU might encounter minor discrepancies.
Fix: Be aware of the minor numerical differences. For most deep learning applications, this level of difference is acceptable. If exact parity is critical, consider using `librosa` for CPU-based processing and carefully benchmark results.
breaking Older versions of `torchlibrosa` (e.g., 0.0.9) had compatibility issues with specific PyTorch versions (e.g., `torch=1.10.0+cu111`), potentially leading to runtime errors related to internal function calls or argument mismatches.
Fix: Upgrade to the latest `torchlibrosa` (0.1.0) and ensure compatibility with your PyTorch version. Check GitHub issues for reported incompatibilities if problems persist.

Install

pip install torchlibrosa Install latest version

Imports

Spectrogram
```
from torchlibrosa import Spectrogram
```

LogmelFilterBank

from torchlibrosa import LogmelFilterBank

STFT
```
from torchlibrosa import STFT
```
ISTFT
```
from torchlibrosa import ISTFT
```
torchlibrosa as tl
```
import torchlibrosa as tl
```
Common alias for convenience.

Quickstart

This quickstart demonstrates how to extract log mel spectrograms from a batch of audio signals using `torchlibrosa`'s `Spectrogram` and `LogmelFilterBank` modules, designed to work seamlessly within a PyTorch `nn.Sequential` model.

import torch
import torchlibrosa as tl

batch_size = 16
sample_rate = 22050
win_length = 2048
hop_length = 512
n_mels = 128

# Create a batch of dummy audio (e.g., for GPU processing)
batch_audio = torch.empty(batch_size, sample_rate).uniform_(-1, 1)

# Instantiate a feature extractor using Sequential for a pipeline
feature_extractor = torch.nn.Sequential(
    tl.Spectrogram(
        hop_length=hop_length,
        win_length=win_length,
    ),
    tl.LogmelFilterBank(
        sr=sample_rate,
        n_mels=n_mels,
        is_log=False, # Default is true
    )
)

# Process the audio to get log mel spectrograms
batch_feature = feature_extractor(batch_audio)

print(f"Input audio shape: {batch_audio.shape}")
print(f"Output feature shape (batch_size, 1, time_steps, mel_bins): {batch_feature.shape}")

view raw JSON →