MixCycle: Unsupervised Speech Separation via Cyclic Mixture Permutation Invariant Training

This repository contains the audio samples and the source code that accompany the paper, available on arXiv and IEEE Xplore.

Audio samples

We provide audio samples to demonstrate the results of the MixCycle method on two different datasets: LibriMix and REAL-M.

Audio samples from the baseline methods on LibriMix are also included: PIT-DM and MixIT.

Note: The audio sample pages were exported to HTML (under docs/) from Jupyter notebooks and are served via GitHub Pages for convenient access directly in the browser. The notebooks/utils.py module contains the helper code that was used by those notebooks.

Source code

We provide the source code under the src directory for reproducibility.

Running the experiments

Prepare the datasets

• LibriMix: GitHub
• REAL-M: Download

Create the environment

Install Anaconda (or the lighter Miniconda) and run the following command:

$ conda env create -f environment.yml

See more info on how to manage conda environments.

Activate the environment

$ conda activate mixcycle

Run the experiments

$ cd src
$ python experiment.py --librimix-root ~/datasets/librimix --exp-root ~/experiments --run librimix_irm
$ python experiment.py --librimix-root ~/datasets/librimix --exp-root ~/experiments --run librimix_5p
$ python experiment.py --librimix-root ~/datasets/librimix --exp-root ~/experiments --run librimix_100p
$ python experiment.py --librimix-root ~/datasets/librimix --realm-root ~/datasets/REAL-M-v0.1.0 --exp-root ~/experiments --run realm

Optionally, you can monitor the training process with TensorBoard by running:

$ tensorboard --logdir ~/experiments

Attribution

The experiments and audio samples in this repository use the following datasets:

• LibriMix — Cosentino et al., "LibriMix: An Open-Source Dataset for Generalizable Speech Separation" (2020). The audio samples in this repository are derived from LibriMix's sep_clean task, and therefore from LibriSpeech (Panayotov et al., "Librispeech: An ASR corpus based on public domain audio books", 2015; CC BY 4.0). They have been modified (mixed, separated, and processed) for this work.
• REAL-M — Subakan et al., "REAL-M: Towards Speech Separation on Real Mixtures" (2022).

Citation (BibTeX)

If you find this repository useful, please cite our work:

@article{karamatli2022unsupervised,
  title={MixCycle: Unsupervised Speech Separation via Cyclic Mixture Permutation Invariant Training},
  author={Karamatl{\i}, Ertu{\u{g}} and K{\i}rb{\i}z, Serap},
  journal={IEEE Signal Processing Letters},
  volume={29},
  number={},
  pages={2637-2641},
  year={2022},
  doi={10.1109/LSP.2022.3232276}
}