/encodec

State-of-the-art deep learning based audio codec supporting both mono 24 kHz audio and stereo 48 kHz audio.

Primary LanguagePythonOtherNOASSERTION

EnCodec: High Fidelity Neural Audio Compression

linter badge tests badge

This is the code for the EnCodec neural codec presented in the High Fidelity Neural Audio Compression [abs]. paper. We provide our two multi-bandwidth models:

  • A causal model operating at 24 kHz on monophonic audio trained on a variety of audio data.
  • A non-causal model operating at 48 kHz on stereophonic audio trained on music-only data.

The 24 kHz model can compress to 1.5, 3, 6, 12 or 24 kbps, while the 48 kHz model support 3, 6, 12 and 24 kbps. We also provide a pre-trained language model for each of the models, that can further compress the representation by up to 40% without any further loss of quality.

For reference, we also provide the code for our novel MS-STFT discriminator and the balancer.

Schema representing the structure of Encodec,
    with a convolutional+LSTM encoder, a Residual Vector Quantization in the middle,
    followed by a convolutional+LSTM decoder. A multiscale complex spectrogram discriminator is applied to the output, along with objective reconstruction losses.
    A small transformer model is trained to predict the RVQ output.

Samples

Samples including baselines are provided on our sample page. You can also have a quick demo of what we achieve for 48 kHz music with EnCodec, along with entropy coding, by clicking the thumbnail (original tracks provided by Lucille Crew and Voyageur I).

Thumbnail for the sample video.
	You will first here the ground truth, then ~3kbps, then 12kbps, for two songs.

What's up?

See the changelog for details on releases.

Installation

EnCodec requires Python 3.8, and a reasonably recent version of PyTorch (1.11.0 ideally). To install EnCodec, you can run from this repository:

pip install -U encodec  # stable release
pip install -U git+https://git@github.com/facebookresearch/encodec#egg=encodec  # bleeding edge
# of if you cloned the repo locally
pip install .

Supported platforms: we officially support only Mac OS X (you might need XCode installed if running on a non Intel Mac), and recent versions of mainstream Linux distributions. We will try to help out on Windows but cannot provide strong support. Any other platform (iOS / Android / onboard ARM) are not supported.

Usage

You can then use the EnCodec command, either as

python3 -m encodec [...]
# or
encodec [...]

If you want to directly use the compression API, checkout encodec.compress and encodec.model. See hereafter for instructions on how to extract the discrete representation.

Model storage

The models will be automatically downloaded on first use using Torch Hub. For more information on where those models are stored, or how to customize the storage location, checkout their documentation.

Compression

encodec [-b TARGET_BANDWIDTH] [-f] [--hq] [--lm] INPUT_FILE [OUTPUT_FILE]

Given any audio file supported by torchaudio on your platform, compresses it with EnCodec to the target bandwidth (default is 6 kbps, can be either 1.5, 3, 6, 12 or 24). OUTPUT_FILE must end in .ecdc. If not provided it will be the same as INPUT_FILE, replacing the extension with .ecdc. In order to use the model operating at 48 kHz on stereophonic audio, use the --hq flag. The -f flag is used to force overwrite an existing output file. Use the --lm flag to use the pretrained language model with entropy coding (expect it to be much slower).

If the sample rate or number of channels of the input doesn't match that of the model, the command will automatically resample / reduce channels as needed.

Decompression

encodec [-f] [-r] ENCODEC_FILE [OUTPUT_WAV_FILE]

Given a .ecdc file previously generated, this will decode it to the given output wav file. If not provided, the output will default to the input with the .wav extension. Use the -f file to force overwrite the output file (be carefull if compress then decompress, not to overwrite your original file !). Use the -r flag if you experience clipping, this will rescale the output file to avoid it.

Compression + Decompression

encodec [-r] [-b TARGET_BANDWIDTH] [-f] [--hq] [--lm] INPUT_FILE OUTPUT_WAV_FILE

When OUTPUT_WAV_FILE has the .wav extension (as opposed to .ecdc), the encodec command will instead compress and immediately decompress without storing the intermediate .ecdc file.

Extracting discrete representations

The EnCodec model can also be used to extract discrete representations from the audio waveform.

from encodec import EncodecModel
from encodec.utils import convert_audio

import torchaudio
import torch

# Instantiate a pretrained EnCodec model
model = EncodecModel.encodec_model_24khz()
# The number of codebooks used will be determined bythe bandwidth selected.
# E.g. for a bandwidth of 6kbps, `n_q = 8` codebooks are used.
# Supported bandwidths are 1.5kbps (n_q = 2), 3 kbps (n_q = 4), 6 kbps (n_q = 8) and 12 kbps (n_q =16) and 24kbps (n_q=32).
# For the 48 kHz model, only 3, 6, 12, and 24 kbps are supported. The number
# of codebooks for each is half that of the 24 kHz model as the frame rate is twice as much.
model.set_target_bandwidth(6.0)

# Load and pre-process the audio waveform
wav, sr = torchaudio.load("<PATH_TO_AUDIO_FILE>")
wav = convert_audio(wav, sr, model.sample_rate, model.channels)
wav = wav.unsqueeze(0)

# Extract discrete codes from EnCodec
with torch.no_grad():
    encoded_frames = model.encode(wav)
codes = torch.cat([encoded[0] for encoded in encoded_frames], dim=-1)  # [B, n_q, T]

Note that the 48 kHz model processes the audio by chunks of 1 seconds, with an overlap of 1%, and renormalizes the audio to have unit scale. For this model, the output of model.encode(wav) would a list (for each frame of 1 second) of a tuple (codes, scale) with scale a scalar tensor.

Installation for development

This will install the dependencies and a encodec in developer mode (changes to the files will directly reflect), along with the dependencies to run unit tests.

pip install -e '.[dev]'

Test

You can run the unit tests with

make tests

FAQ

Please check this section before opening an issue.

Out of memory errors with long files

We do not try to be smart about long files, and we apply the model at once on the entire file. This can lead to a large memory usage and result in the process being killed. At the moment we will not support this use case.

Bad interactions between DistributedDataParallel and the RVQ code

We do not use DDP, instead we recommend using the routines in encodec/distrib.py, in particular encodec.distrib.sync_buffer and encodec.distrib.sync_grad.

Citation

If you use this code or results in your paper, please cite our work as:

@article{defossez2022highfi,
  title={High Fidelity Neural Audio Compression},
  author={Défossez, Alexandre and Copet, Jade and Synnaeve, Gabriel and Adi, Yossi},
  journal={arXiv preprint arXiv:2210.13438},
  year={2022}
}

License

This repository is released under the CC-BY-NC 4.0. license as found in the LICENSE file.