Onssen, pronounced as おんせん(温泉, Japanese hot spring), is a PyTorch-based library for speech separation, speech enhancement, or speech style transformation.

• Provide template classes for data, model, and evaluation
• Move models to separate folders (i.e. Kaldi style)
• Reproduce scores and upload pretrained models
• Finish inference method for online separation

• Add evaluation method for deep clustering
• Use W_{MR} weight in deep clustering
• Minor changes

• Deep Clustering
• Chimera Net
• Chimera++
• Phase Estimation Network
• Speech Enhancement with Restoration Layers

• Wsj0-2mix (http://www.merl.com/demos/deep-clustering)
• Daps (https://archive.org/details/daps_dataset)
• Edinburgh-TTS (https://datashare.is.ed.ac.uk/handle/10283/2791)

• PyTorch
• LibRosa
• NumPy

You can simply use the existing config JSON file or customize your config file to train the enhancement or separation model. under the egs/wsj0-2mix/deep_clustering/ directory:

python run.py -c config.json

If you use onssen for your research project, please cite one of the following bibtex citations:

@article{ni2019onssen,
title={Onssen: an open-source speech separation and enhancement library},
author={Ni, Zhaoheng and Mandel, Michael I},
journal={arXiv preprint arXiv:1911.00982},
year={2019}
}

@Misc{onssen,
    author = {Zhaoheng Ni and Michael Mandel},
    title = "ONSSEN: An Open-source Speech Separation and Enhancement Library",
    howpublished = {\url{https://github.com/speechLabBcCuny/onssen}},
    year =        {2019}
}