Music Source Separation in the Waveform Domain

Demucs was just updated!: much better SDR, smaller models, more data augmentation and PyPI support.

Windows users: for the moment things will be a bit broken, please stick with the old version for a bit (git checkout v1), while I try to find various fixes.

For the initial version of Demucs: Go this commit. If you are experiencing issues and want the old Demucs back, please fill an issue, and then you can get back to the v1 with git checkout v1.

We provide an implementation of Demucs and Conv-Tasnet for music source separation on the MusDB dataset. They can separate drums, bass and vocals from the rest with state-of-the-art results, surpassing previous waveform or spectrogram based methods. The architecture and results obtained are detailed in our paper Music Source Separation in the waveform domain.

Demucs is based on U-Net convolutional architecture inspired by Wave-U-Net and SING, with GLUs, a BiLSTM between the encoder and decoder, specific initialization of weights and transposed convolutions in the decoder.

Conv-Tasnet is a separation model developed for speech which predicts a mask on a learnt over-complete linear representation using a purely convolutional model with stride of 1 and dilated convolutional blocks. We reused the code from the kaituoxu/Conv-TasNet repository and added support for multiple audio channels.

Demucs achieves a state-of-the-art SDR performance of 6.3 when trained only on MusDB. Conv-Tasnet achieves an SDR of 5.7, to be compared with the best performing spectrogram domain model D3Net with an average SDR of 6. Unlike Conv-Tasnet, Demucs reacts positively to pitch/tempo shift augmentation (+0.5 SDR). However, Demucs still suffers from leakage from other sources, in particular between the vocals and other sources, which is less of a problem for Conv-Tasnet. When trained with 150 extra tracks, Demucs reaches an SDR of 6.8, and even surpasses the IRM oracle for the bass source (7.6 against 7.1 for the oracle). See our paper Section 6 for more details or listen to our audio samples .

Important news if you are already using Demucs

See the release notes for more details.

11/05/2021: Adding support for MusDB-HQ and arbitrary wav set, for the MDX challenge. For more information on joining the challenge with Demucs see the Demucs MDX instructions
28/04/2021: Demucs v2, with extra augmentation and DiffQ based quantization. EVERYTHING WILL BREAK, please restart from scratch following the instructions hereafter. This version also adds overlap between prediction frames, with linear transition from one to the next, which should prevent sudden changes at frame boundaries. Also, Demucs is now on PyPI, so for separation only, installation is as easy as pip install demucs :)
13/04/2020: Demucs released under MIT: We are happy to release Demucs under the MIT licence. We hope that this will broaden the impact of this research to new applications.

Comparison with other models

An audio comparison of Demucs and Conv-Tasnet with other state-of-the-art methods such as Wave-U-Net, OpenUnmix or MMDenseLSTM is available on the audio comparison page. We provide hereafter a summary of the different metrics presented in the paper. You can also compare Spleeter, Open-Unmix, Demucs and Conv-Tasnet on one of my favorite songs on our soundcloud playlist.

Comparison of accuracy

Overall SDR is the mean of the SDR for each of the 4 sources, MOS Quality is a rating from 1 to 5 of the naturalness and absence of artifacts given by human listeners (5 = no artifacts), MOS Contamination is a rating from 1 to 5 with 5 being zero contamination by other sources. We refer the reader to our paper, Section 5 and 6, for more details.

Model	Domain	Extra data?	Overall SDR	MOS Quality	MOS Contamination
Open-Unmix	spectrogram	no	5.3	3.0	3.3
D3Net	spectrogram	no	6.0	-	-
Wave-U-Net	waveform	no	3.2	-	-
Demucs (this)	waveform	no	6.3	3.2	3.3
Conv-Tasnet (this)	waveform	no	5.7	2.9	3.4
Demucs (this)	waveform	150 songs	6.8	-	-
Conv-Tasnet (this)	waveform	150 songs	6.3	-	-
MMDenseLSTM	spectrogram	804 songs	6.0	-	-
D3Net	spectrogram	1.5k songs	6.7	-	-
Spleeter	spectrogram	25k songs	5.9	-	-

Requirements

You will need at least Python 3.7. See requirements.txt for requirements for separation only, and environment-[cpu|cuda].yml if you want to train a new model.

For Windows users

Everytime you see python3, replace it with python.exe. You should always run commands from the Anaconda console.

For musicians

If you just want to use Demucs to separate tracks, you can install it with

python3 -m pip -U install demucs

Advanced OS support are provided on the following page, you must read the page for your OS before posting an issues:

If you are using Windows: Windows support.
If you are using MAC OS X: Mac OS X support.
If you are using Linux: Linux support.

For machine learning scientists

If you have anaconda installed, you can run from the root of this repository:

conda env update -f environment-cpu.yml # if you don't have GPUs
conda env update -f environment-cuda.yml # if you have GPUs
conda activate demucs
pip install -e .

This will create a demucs environment with all the dependencies installed.

You will also need to install soundstretch/soundtouch: on Mac OSX you can do brew install sound-touch, and on Ubuntu sudo apt-get install soundstretch. This is used for the pitch/tempo augmentation.

Running in Docker

Thanks to @xserrat, there is now a Docker image definition ready for using Demucs. This can ensure all libraries are correctly installed without interfering with the host OS. See his repo Docker Facebook Demucs for more information.

Running from Colab

I made a Colab to easily separate track with Demucs. Note that transfer speeds with Colab are a bit slow for large media files, but it will allow you to use Demucs without installing anything.

Demucs on Google Colab

Separating tracks

In order to try Demucs or Conv-Tasnet on your tracks, simply run from the root of this repository

python3 -m demucs.separate PATH_TO_AUDIO_FILE_1 [PATH_TO_AUDIO_FILE_2 ...] # for Demucs
python3 -m demucs.separate --mp3 PATH_TO_AUDIO_FILE_1 --mp3-bitrate BITRATE # output files saved as MP3
python3 -m demucs.separate -n tasnet PATH_TO_AUDIO_FILE_1 ... # for Conv-Tasnet

If you have a GPU, but you run out of memory, please add -d cpu to the command line. See the section hereafter for more details on the memory requirements for GPU acceleration.

Separated tracks are stored in the separated/MODEL_NAME/TRACK_NAME folder. There you will find four stereo wav files sampled at 44.1 kHz: drums.wav, bass.wav, other.wav, vocals.wav (or .mp3 if you used the --mp3 option).

All audio formats supported by torchaudio can be processed (i.e. wav, mp3, flac, ogg/vorbis etc.). Audio is resampled on the fly if necessary. The output will be a wave file, either in int16 format or float32 (if --float32 is passed). You can pass --mp3 to save as mp3 instead, and set the bitrate with --mp3-bitrate (default is 320kbps).

Other pre-trained models can be selected with the -n flag. The list of pre-trained models is:

demucs: Demucs trained on MusDB,
demucs_quantized: Quantized Demucs with diffq, this is much smaller (150MB instead of 1GB) and quality should be exactly the same. Let me know if you disagree. As a result, this is the one used by default.
demucs_extra: Demucs trained with extra training data,
demucs48_hq: Demucs with 48 initial hidden channels, trained on MusDB-HQ, used as a baseline for the Music Demixing Challenge 2021,
tasnet: Conv-Tasnet trained on MusDB,
tasnet_extra: Conv-Tasnet trained with extra training data.

The --shifts=SHIFTS performs multiple predictions with random shifts (a.k.a the shift trick) of the input and average them. This makes prediction SHIFTS times slower but improves the accuracy of Demucs by 0.2 points of SDR. It has limited impact on Conv-Tasnet as the model is by nature almost time equivariant. The value of 10 was used on the original paper, although 5 yields mostly the same gain. It is deactivated by default but it does make vocals a bit smoother.

The --overlap option controls the amount of overlap between prediction windows (for Demucs one window is 10 seconds). Default is 0.25 (i.e. 25%) which is probably fine.

Memory requirements for GPU acceleration

If you want to use GPU acceleration, you will need at least 8GB of RAM on your GPU for demucs and 4GB for tasnet. Sorry, the code for demucs is not super optimized for memory! If you do not have enough memory on your GPU, simply add -d cpu to the command line to use the CPU. With Demucs, processing time should be roughly equal to the duration of the track.

Examining the results from the paper experiments

The metrics for our experiments are stored in the results folder. In particular museval json evaluations are stored in results/evals/EXPERIMENT NAME/results. You can aggregate and display the results using

python3 valid_table.py -p # show valid loss, aggregated with multiple random seeds
python3 result_table.py -p # show SDR on test set, aggregated with multiple random seeds
python3 result_table.py -p SIR # also SAR, ISR, show other metrics

The std column shows the standard deviation divided by the square root of the number of runs.

Training Demucs and evaluating on the MusDB dataset

If you want to train Demucs from scratch, you will need a copy of the MusDB dataset. It can be obtained on the MusDB website. To start training on a single GPU or CPU, use:

python3 -m demucs -b 4  --musdb MUSDB_PATH # Demucs
python3 -m demucs -b 4  --musdb MUSDB_PATH --tasnet --samples=80000 --split_valid # Conv-Tasnet

The -b 4 flag will set the batch size to 4. The default is 4 and will crash on a single GPU. Demucs was trained on 8 V100 with 32GB of RAM. The default parameters (batch size, number of channels etc) might not be suitable for 16GB GPUs. To train on all available GPUs, use:

python3 run.py --musdb MUSDB_PATH [EXTRA_FLAGS]

This will launch one process per GPU and report the output of the first one. When interrupting such a run, it is possible some of the children processes are not killed properly, be mindful of that. If you want to use only some of the available GPUs, export the CUDA_VISIBLE_DEVICES variable to select those.

To see all the possible options, use python3 -m demucs --help.

MusDB HQ

To train on MusDB HQ, use the following flags:

python3 -m demucs -b 4 --musdb MUSDB_HQ_PATH --is_wav [...]

Custom wav dataset

You can trained on a custom wav dataset using the following command. At the moment, you still need to pass the MusDB path for evaluation, and the model must use the standard sources (bass, drums, other, vocals). However, it should be relatively easy to fork the code to support different patterns.

python3 -m demucs -b 4 --wav PATH_TO_WAV_DATASET [...]

The folder PATH_TO_WAV_DATASET should contain two sub-directories : train and valid. Each of those should contain one folder per track. Each track folder must contain one file for each source (drums.wav, bass.wav, other.wav, vocals.wav) and one file for the mixture (mixture.wav).

Fine tuning

You can fine tune from one of the pre-trained models listed in the Separating tracks Section by passing the --init=PRETRAINED_NAME, i.e. for Demucs or ConvTasnet:

python3 -m demucs -b 4  --musdb MUSDB_PATH --init demucs # Demucs
python3 -m demucs -b 4  --musdb MUSDB_PATH --tasnet --samples=80000 --split_valid --init tasnet # Conv-Tasnet

About checkpointing

Demucs will automatically generate an experiment name from the command line flags you provided. It will checkpoint after every epoch. If a checkpoint already exist for the combination of flags you provided, it will be automatically used. In order to ignore/delete a previous checkpoint, run with the -R flag. The optimizer state, the latest model and the best model on valid are stored. At the end of each epoch, the checkpoint will erase the one from the previous epoch. By default, checkpoints are stored in the ./checkpoints folder. This can be changed using the --checkpoints CHECKPOINT_FOLDER flag.

Not all options will impact the name of the experiment. For instance --workers is not shown in the name, therefore, changing this parameter will not impact the checkpoint file used. Refer to parser.py for more details.

Test set evaluations

Test set evaluations computed with museval will be stored under evals/EXPERIMENT NAME/results. The experiment name is the first thing printed when running python3 run.py or python3 -m demucs. If you used the flag --save, there will also be a folder evals/EXPERIMENT NAME/wavs containing all the extracted waveforms.

Running on a cluster

If you have a cluster available with Slurm, you can set the run_slurm.py as the target of a slurm job, using as many nodes as you want and a single task per node. run_slurm.py will create one process per GPU and run in a distributed manner. Multinode training is supported.

Extracting Raw audio for faster loading

We observed that loading from compressed mp4 audio lead to unreliable speed, sometimes reducing by a factor of 2 the number of iterations per second. It is possible to extract all data to raw PCM f32e format. If you wish to store the raw data under RAW_PATH, run the following command first:

python3 -m demucs.raw [--workers=10] MUSDB_PATH RAW_PATH

You can then train using the --raw RAW_PATH flag, for instance:

python3 run.py --raw RAW_PATH --musdb MUSDB_PATH

You still need to provide the path to the MusDB dataset as we always load the test set from the original MusDB.

Results reproduction

To reproduce the performance of the main Demucs model in our paper:

# Extract raw waveforms. This is optional
python3 -m demucs.data MUSDB_PATH RAW_PATH
export DEMUCS_RAW=RAW_PATH
# Train models with default parameters and multiple seeds
python3 run.py --seed 42 # for Demucs
python3 run.py --seed 42 --tasnet --X=10 --samples=80000 --epochs=180 --split_valid # for Conv-Tasnet
# Repeat for --seed = 43, 44, 45 and 46

You can visualize the results aggregated on multiple seeds using

python3 valid_table.py # compare validation losses
python3 result_table.py # compare test SDR
python3 result_table.py SIR # compare test SIR, also available ISR, and SAR

You can look at our exploration file dora.py to see the exact flags for all experiments (grid search and ablation study). If you have a Slurm cluster, you can also try adapting it to run on your own.

Environment variables

If you do not want to always specify the path to MUSDB, you can export the following variables:

export DEMUCS_MUSDB=PATH TO MUSDB
# Optionally, if you extracted raw pcm data
# export DEMUCS_RAW=PATH TO RAW PCM

How to cite

@article{defossez2019music,
  title={Music Source Separation in the Waveform Domain},
  author={D{\'e}fossez, Alexandre and Usunier, Nicolas and Bottou, L{\'e}on and Bach, Francis},
  journal={arXiv preprint arXiv:1911.13254},
  year={2019}
}

License

Demucs is released under the MIT license as found in the LICENSE file.

The file demucs/tasnet.py is adapted from the kaituoxu/Conv-TasNet repository. It was originally released under the MIT License updated to support multiple audio channels.

LearnNLP/demucs