Improving DNN-based Music Source Separation using Phase Features

This repository contains the implementation of Improving DNN-based Music Source Separation using Phase Features, from Muth et al.

Preparation

Clone this repository

git clone https://github.com/aRI0U/music-source-separation

Install all Python dependencies

cd music-source-separation
pip install -r requirements.txt

Download the MUSDB18 dataset. Access on request here.
Put the archive containing MUSDB18 in a folder named datasets and extract it with the provided script.

mkdir datasets
mv /path/to/musdb18.zip datasets
./utils/setup_musdb.sh

The architecture of the model contains a bias and scale layers which are initialized respectively with the mean and standard deviation per frequency bin over the training set. In order to compute these statistics, run the following:

python3 utils/compute_stats.py

Warning: This script computes the STFT of the tracks in the dataset. The parameters of the STFT given by default are the same as the default options of train.py. If you want to change the STFT options in train.py, you will have to change the parameters of this compute_stats.py consequently.

Usage

Train a model
```
python3 train.py
```
A lot of options can be configured through command-line arguments. Type python3 train.py --help for an exhaustive detailed list of those options.

Folder structure

When you run train.py, a new folder is created in folder runs. You can specify the name of the experiment with option --name. Otherwise a timestamp will be used.

Each folder of runs contains three subfolders:

A checkpoints directory containing the saved models along training.
A logs directory containing the Tensorboard logs. One can thus track the evolution of the loss along the training by running on a separate terminal
```
tensorboard --logdir runs
```
Samples of separed audio from the validation set can also be listened along training with Tensorboard.
A results directory containing signals separed by the model.

Since saving checkpoints and audio extracts is a quite slow operation, it is done only every 5 epochs. You can change this behaviour with option --save_freq.

Citation

This work implements the model presented in Improving DNN-based Music Source Separation using Phase Features. To our knowledge, no official implementation of this paper is available.

To cite the original paper:

@article{Muth18,
  archivePrefix = {arXiv},
  arxivId       = {1807.02710},
  author        = {Muth, Joachim and
                   Uhlich, Stefan and
                   Perraudin, Nathanael and
                   Kemp, Thomas and
                   Cardinaux, Fabien and
                   Mitsufuji, Yuki},
  eprint        = {1807.02710},
  title         = {{Improving DNN-based Music Source Separation using Phase Features}},
  url           = {http://arxiv.org/abs/1807.02710},
  year          = {2018}
}

A part of the code of this repository (the major part of data.py and maybe a few lines somewhere else) come from Open-Unmix.

For all experiments, the model of this repository uses the MUSDB18 dataset.

@misc{musdb18,
  author       = {Rafii, Zafar and
                  Liutkus, Antoine and
                  Fabian-Robert St{\"o}ter and
                  Mimilakis, Stylianos Ioannis and
                  Bittner, Rachel},
  title        = {The {MUSDB18} corpus for music separation},
  month        = dec,
  year         = 2017,
  doi          = {10.5281/zenodo.1117372},
  url          = {https://doi.org/10.5281/zenodo.1117372}
}

Abefocuson/music-source-separation

Improving DNN-based Music Source Separation using Phase Features

Preparation

Usage

Folder structure

Citation