torchaudio: an audio library for PyTorch
The aim of torchaudio is to apply PyTorch to the audio domain. By supporting PyTorch, torchaudio follows the same philosophy of providing strong GPU acceleration, having a focus on trainable features through the autograd system, and having consistent style (tensor names and dimension names). Therefore, it is primarily a machine learning library and not a general signal processing library. The benefits of PyTorch can be seen in torchaudio through having all the computations be through PyTorch operations which makes it easy to use and feel like a natural extension.
- Support audio I/O (Load files, Save files)
- Load the following formats into a torch Tensor using SoX
- mp3, wav, aac, ogg, flac, avr, cdda, cvs/vms,
- aiff, au, amr, mp2, mp4, ac3, avi, wmv,
- mpeg, ircam and any other format supported by libsox.
- Kaldi (ark/scp)
- Load the following formats into a torch Tensor using SoX
- Dataloaders for common audio datasets (VCTK, YesNo)
- Common audio transforms
- Compliance interfaces: Run code using PyTorch that align with other libraries
- PyTorch (See below for the compatible versions)
- libsox v14.3.2 or above (only required when building from source)
- [optional] vesis84/kaldi-io-for-python commit cb46cb1f44318a5d04d4941cf39084c5b021241e or above
The following are the corresponding
torchaudio versions and supported Python versions.
To install the latest version using anaconda, run:
conda install -c pytorch torchaudio
To install the latest pip wheels, run:
pip install torchaudio -f https://download.pytorch.org/whl/torch_stable.html
(If you do not have torch already installed, this will default to installing torch from PyPI. If you need a different torch configuration, preinstall torch before running this command.)
Note that nightly build is build on PyTorch’s nightly build. Therefore, you need to install the latest PyTorch when you use nightly build of torchaudio.
pip install numpy pip install --pre torchaudio -f https://download.pytorch.org/whl/nightly/torch_nightly.html
conda install -y -c pytorch-nightly torchaudio
If your system configuration is not among the supported configurations above, you can build torchaudio from source.
This will require libsox v14.3.2 or above.
Click here for the examples on how to install SoX
brew install sox
sudo apt-get install sox libsox-dev libsox-fmt-all
conda install -c conda-forge sox
# Linux python setup.py install # OSX MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install
Alternatively, the build process can build libsox and some optional codecs statically and torchaudio can link them, by setting environment variable
BUILD_SOX=1. The build process will fetch and build libmad, lame, flac, vorbis, opus, and libsox before building extension. This process requires
# Linux BUILD_SOX=1 python setup.py install # OSX BUILD_SOX=1 MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install
This is known to work on linux and unix distributions such as Ubuntu and CentOS 7 and macOS. If you try this on a new system and find a solution to make it work, feel free to share it by opening an issue.
checking build system type… ./config.guess: unable to guess system type
Since the configuration file for codecs are old, they cannot correctly detect the new environments, such as Jetson Aarch. You need to replace the
config.guess file in
./third_party/tmp/libmad-0.15.1b/config.guess with the latest one.
See also: #658
Undefined reference to `tgetnum’ when using `BUILD_SOX`
If while building from within an anaconda environment you come across errors similar to the following:
../bin/ld: console.c:(.text+0xc1): undefined reference to `tgetnum'
conda-forge before running
python setup.py install:
# Install ncurses from conda-forge conda install -c conda-forge ncurses
import torchaudio waveform, sample_rate = torchaudio.load('foo.wav') # load tensor from file torchaudio.save('foo_save.wav', waveform, sample_rate) # save tensor to file
import torchaudio torchaudio.set_audio_backend("soundfile") # switch backend waveform, sample_rate = torchaudio.load('foo.wav') # load tensor from file, as usual torchaudio.save('foo_save.wav', waveform, sample_rate) # save tensor to file, as usual
Unlike SoX, SoundFile does not currently support mp3.
API Reference is located here: http://pytorch.org/audio/
With torchaudio being a machine learning library and built on top of PyTorch, torchaudio is standardized around the following naming conventions. Tensors are assumed to have “channel” as the first dimension and time as the last dimension (when applicable). This makes it consistent with PyTorch’s dimensions. For size names, the prefix
n_ is used (e.g. “a tensor of size (
n_mel)”) whereas dimension names do not have this prefix (e.g. “a tensor of dimension (channel, time)”)
waveform: a tensor of audio samples with dimensions (channel, time)
sample_rate: the rate of audio dimensions (samples per second)
specgram: a tensor of spectrogram with dimensions (channel, freq, time)
mel_specgram: a mel spectrogram with dimensions (channel, mel, time)
hop_length: the number of samples between the starts of consecutive frames
n_fft: the number of Fourier bins
n_mfcc: the number of mel and MFCC bins
n_freq: the number of bins in a linear spectrogram
min_freq: the lowest frequency of the lowest band in a spectrogram
max_freq: the highest frequency of the highest band in a spectrogram
win_length: the length of the STFT window
window_fn: for functions that creates windows e.g.
Transforms expect and return the following dimensions.
Spectrogram: (channel, time) -> (channel, freq, time)
AmplitudeToDB: (channel, freq, time) -> (channel, freq, time)
MelScale: (channel, freq, time) -> (channel, mel, time)
MelSpectrogram: (channel, time) -> (channel, mel, time)
MFCC: (channel, time) -> (channel, mfcc, time)
MuLawEncode: (channel, time) -> (channel, time)
MuLawDecode: (channel, time) -> (channel, time)
Resample: (channel, time) -> (channel, time)
Fade: (channel, time) -> (channel, time)
Vol: (channel, time) -> (channel, time)
Complex numbers are supported via tensors of dimension (…, 2), and torchaudio provides
angle to convert such a tensor into its magnitude and phase. Here, and in the documentation, we use an ellipsis “…” as a placeholder for the rest of the dimensions of a tensor, e.g. optional batching and channel dimensions.
Please refer to CONTRIBUTING.md
Disclaimer on Datasets
This is a utility library that downloads and prepares public datasets. We do not host or distribute these datasets, vouch for their quality or fairness, or claim that you have license to use the dataset. It is your responsibility to determine whether you have permission to use the dataset under the dataset’s license.
If you’re a dataset owner and wish to update any part of it (description, citation, etc.), or do not want your dataset to be included in this library, please get in touch through a GitHub issue. Thanks for your contribution to the ML community!