Unsupervised Audio Source Separation using Generative Priors

Refer to diagram below for a summary of the approach. Here is a demo (please view in Firefox for best WAV file support).

Requirements

• python 3.6
• numpy >= 1.11.0
• tensorflow = 1.12.0
• scikit-learn >= 0.18
• matplotlib >= 2.1.0
• librosa

Prerequisites

The Digit, Drum and Piano audio datasets can be downloaded from: https://github.com/chrisdonahue/wavegan

The datasets need to be in the directory /data

Data Processing

Run create_dataset.py with appropriate arguments to randomly sample 1000 examples of digit, drums and piano audio from the respective datasets and form the normalized mixtures.

Pre-trained GAN Priors

Download the Pre-trained WaveGAN Priors from here and save the contents in the respective folders in the /ckpts directory

Perform Source Separation

Run run_twosourcesep.py to perform two source separation on the appropriate mixture combination (digit-drums, digit-piano or drums-piano).

To specify the paths for the pre-trained models, mixture data, results, type of sources and other hyperparameters,modify the arguments of run_twosourcesep.py

Run run_threesourcesep.py to perform three source separation on the digit-drums-piano mixture combination

To specify the paths for the pre-trained models, mixture data, results, type of sources and other hyperparameters,modify the arguments of run_threesourcesep.py

Citations

Our paper is cited as:

@article{narayanaswamy2020,
  title={Unsupervised Audio Source Separation using Generative Priors},
  author={Vivek Narayanaswamy, Jayaraman J. Thiagarajan, Rushil Anirudh and Andreas Spanias},
  journal={arXiv preprint arXiv:2005.13769},
  year={2020}
}