This repository provides a PyTorch implementation of "Semi-supervised Monaural Singing Voice Separation with a Masking Network Trained on Synthetic Mixtures" (paper).

The network learns to seperate vocals from music, by training on a set of samples of mixed music (singing and instrumental) and an unmatched set of instrumental music. Comparison with a fully supervised method can be found here

The algorithm is based on two main ideas:

1. Utilizing GAN losses for alinging the distribution of masked samples with instrumental music and the distribution of synthetic samples with real mixture samples.
2. Reconstruction losses based on the architecture and the superposition property of audio channels, creating more stable constrain than the GAN loss.

A conda environment file is available in the repository.

• Python 3.6 +
• Pytorch 0.4.0
• Torchvision
• Tensorboardx
• librosa
• tqdm
• imageio
• opencv
• musdb
• dsd100

$ git clone https://github.com/sagiebenaim/Singing.git
$ cd Singing/

For creating and activating the conda environment:

$ conda env create -f environment.yml
$ source activate singing

To download the DSD100 dataset:

$ bash download.sh DSD100

To download the MUSDB18 dataset, you must request access to the dataset from this website. Then, you need unzip the file to ./dataset/ folder

For creating the spectrograms for training, run the following command

$ python dataset.py --dataset DSD100 

The default spectrograms are for vocals and accompaniment seperation. for drums/bass, use the --target drums/--target bass option

Run the training script below.

$ python train.py --config configs/vocals_new.yaml

To test on the DSD100 dataset:

$ python test.py

To separate music channels (default:vocals and accompaniment) for DSD100, using the pretrained model, run the script below. The audio files will be saved into ./outputs directory.

$ python create_songs.py

The script contains the function create_songs for creating vocals and accompaniment for any given song directory, for file format supported by soundfile. for doing so, create a folder named ./input, move there the files and run the script:

$ python create_songs.py --input custom

If you found this code useful, please cite the following paper:

  @inproceedings{michelashvili2018singing,
  title={Semi-Supervised Monaural Singing Voice Separation With a Masking Network Trained on Synthetic Mixtures},
  author={Michael Michelashvili and Sagie Benaim and Lior Wolf},
  booktitle={ICASSP},
  year={2019},
}

This repository is based on the code from MUNIT