TorchAudio

2.11.0 · active · verified Sun Apr 05

TorchAudio is an open-source library for audio and signal processing with PyTorch, providing functions, datasets, model implementations, and application components for machine learning tasks. While it has transitioned into a maintenance phase since version 2.8/2.9 to reduce redundancies and focus on ML audio processing, it continues to release new versions (currently 2.11.0) in alignment with PyTorch releases.

Warnings

breaking Breaking API changes: Most APIs explicitly marked as 'drop' were deprecated in TorchAudio 2.8 and subsequently removed in 2.9. This can cause `AttributeError` or `ImportError` if upgrading from older versions.
Fix: Refer to the TorchAudio 2.9 migration guide on the official documentation to identify and update removed APIs. Many functions were consolidated or moved to `torchcodec`.
gotcha The `torchaudio.load()` and `torchaudio.save()` functions (since 2.9) now internally rely on the `torchcodec` library. While they maintain a compatible API, some parameters like `normalize`, `buffer_size`, and `backend` are ignored. For optimal performance and full control, it is recommended to directly use `torchcodec.decoders.AudioDecoder` and `torchcodec.encoders.AudioEncoder`.
Fix: Review calls to `torchaudio.load()` and `torchaudio.save()`. Migrate to native `torchcodec` APIs if you rely on parameters that are now ignored or for improved performance. Ensure `torchcodec` is installed and compatible with your `torch` version.
breaking Strict PyTorch Version Compatibility: TorchAudio releases are tightly coupled with specific PyTorch versions. Using mismatched versions of `torch` and `torchaudio` will lead to runtime errors, particularly with C++ extensions.
Fix: Always install `torch` and `torchaudio` versions that are explicitly listed as compatible in the official TorchAudio compatibility matrix (e.g., `torch==X.Y.Z` and `torchaudio==X.Y.Z`).
gotcha FFmpeg dependency for I/O: TorchAudio's audio loading and saving functionalities, particularly through the `torchcodec` backend, heavily rely on FFmpeg being installed and accessible on your system. Missing or incompatible FFmpeg versions can lead to `RuntimeError` during audio processing.
Fix: Ensure FFmpeg is installed and discoverable by your system. For `conda` environments, `conda install -c conda-forge 'ffmpeg<8'` (or a suitable version) is often effective. Refer to `torchcodec` installation instructions for detailed FFmpeg compatibility.

Install

pip install torchaudio Default Install

Imports

torchaudio
```
import torchaudio
```
transforms
```
from torchaudio import transforms
```
functional
```
from torchaudio import functional
```

Quickstart

This quickstart demonstrates how to import TorchAudio, create a dummy audio waveform, and apply a common audio transformation like MelSpectrogram. In practice, `torchaudio.load` is used to load actual audio files.

import torch
import torchaudio
from torchaudio import transforms

# Create a dummy waveform (1 channel, 16000 samples at 16kHz)
# In a real scenario, you would load an audio file: waveform, sample_rate = torchaudio.load("path/to/audio.wav")
waveform = torch.randn(1, 16000)
sample_rate = 16000

# Define a MelSpectrogram transform
melspectrogram_transform = transforms.MelSpectrogram(sample_rate=sample_rate, n_mels=128)

# Apply the transform
melspectrogram = melspectrogram_transform(waveform)

print(f"Waveform shape: {waveform.shape}")
print(f"MelSpectrogram shape: {melspectrogram.shape}")
# Expected output: 
# Waveform shape: torch.Size([1, 16000])
# MelSpectrogram shape: torch.Size([1, 128, X]) where X depends on n_fft and hop_length

view raw JSON →