Skip to content

Latest commit

 

History

History
72 lines (54 loc) · 4.02 KB

README.md

File metadata and controls

72 lines (54 loc) · 4.02 KB
Raw

Unsupervised Interpretable Representation Learning for Singing Voice Separation

This repository contains the PyTorch (1.4) implementation of our method for representation learning. Our method is based on (convolutional) neural networks, to learn representations from music signals that could be used for singing voice separation. The trick here is that the proposed method is employing cosine functions at the decoding stage. The resulting representation is non-negative and real-valued, and it could employed, fairly easily, by current supervised models for music source separation. The proposed method is inspired by Sinc-Net and dDSP.

Authors

S.I. Mimilakis, K. Drossos, G. Schuller

What's inside?

  • Code for the neural architectures used in our study and their corresponding minimization objectives (nn_modules/)
  • Code for performing the unsupervised training (scripts/exp_rl_*)
  • Code for reconstructing the signal(s) (scripts/exp_fe_test.py)
  • Code for inspecting the outcome(s) of the training (scripts/make_plots.py)
  • Code for visualizing loss functions, reading/writing audio files, and creating batches (tools/)
  • Perks (unreported implementations/routines)
    • The discriminator-like objective, as a proxy to mutual information, reported here

What's not inside!

How to use

Training

  1. Download the dataset and declare the path of the downloaded dataset in tools/helpers.py
  2. Apply any desired changes to the model by tweeking the parameters in settings/rl_experiment_settings.py
  3. Execute scripts/exp_rl_vanilla.py for the basic method
  4. Execute scripts/exp_rl_sinkhorn.py for the extended method, using Sinkhorn distances

Testing

  1. Download the dataset and declare the path of the downloaded dataset in tools/helpers.py
  2. Download the results and place them under the results folder
  3. Load up the desired model by declaring the experiment id in settings/rl_experiment_settings.py (e.g. r-mcos8)
  4. Execute scripts/exp_fe_test.py (some arguments for plotting and file writing are necesary)

Reference

If you find this code useful for your research, cite our papers:

  @inproceedings{mim20_uirl_eusipco,  
  author={S. I. Mimilakis and K. Drossos and G. Schuller},  
  title={Unsupervised Interpretable Representation Learning for Singing Voice Separation},  
  year={2020},
  booktitle={Proceedings of the 27th European Signal Processing Conference (EUSIPCO 2020)}
  }
  @misc{mimilakis2020revisiting,
  title={Revisiting Representation Learning for Singing Voice Separation with Sinkhorn Distances},
  author={S. I. Mimilakis and K. Drossos and G. Schuller},
  year={2020},
  eprint={2007.02780},
  archivePrefix={arXiv},
  primaryClass={cs.SD}
  }

Acknowledgements

Stylianos Ioannis Mimilakis is supported in part by the German Research Foundation (AB 675/2-1, MU 2686/11-1).

License

MIT