Docker Facebook Demucs

This repository dockerizes Demucs to split music tracks into different tracks (bass, drums, voice, others).

Usage

Clone this repository

git clone https://github.com/xserrat/docker-facebook-demucs.git demucs

Split a music track

1. Copy the track you want to split into the input folder (e.g., input/mysong.mp3).
2. Execute demucs via the run job in the Makefile, specifying the track argument with only the name of the file:

make run track=mysong.mp3

This process will take some time the first time it is run, as the execution will:

• Download the Docker image that is setup to run the facebook demucs script.
• Download the pretrained models.
• Execute demucs to split the track.

Subsequent runs will not need to download the Docker image or download the models, unless the model specified has not yet been used.

Options

The following options are available when splitting music tracks with the run job:

Option	Default Value	Description
gpu	false	Enable Nvidia CUDA support (requires an Nvidia GPU).
model	demucs	The model used for audio separation. See https://github.com/facebookresearch/demucs#separating-tracks for a list of available models to use.
mp3output	false	Output separated audio in mp3 format instead of the default wav format.
shifts	1	Perform multiple predictions with random shifts (a.k.a the shift trick) of the input and average them. This makes prediction SHIFTS times slower. Don't use it unless you have a GPU.
overlap	0.25	Control the amount of overlap between prediction windows. Default is 0.25 (i.e. 25%) which is probably fine. It can probably be reduced to 0.1 to improve separation speed.
jobs	1	Specify the number of parallel jobs to run during separation. This will multiply the amount of RAM used by the same number, so be careful!
splittrack		Individual track to split/separate from the others (e.g., you only want to separate drums). Valid options are bass, drums, vocals and other. Other values may be allowed if the model can separate additional track types.

Example commands:

# Use the "fine tuned" demucs model
make run track=mysong.mp3 model=htdemucs_ft

# Enable Nvidia CUDA support and output separated audio in mp3 format
make run track=mysong.mp3 gpu=true mp3output=true

Run Interactively

To experiment with other demucs options on the command line, you can also run the Docker image interactively via the run-interactive job. Note that only the gpu option is applicable for this job.

Example:

make run-interactive gpu=true

Building the Image

The Docker image can be built locally via the build job:

make build

License

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