https://github.com/desh2608/css

PyTorch implementation of Continuous Speech Separation
https://github.com/desh2608/css

Last synced: about 2 months ago
JSON representation

PyTorch implementation of Continuous Speech Separation

• Host: GitHub
• URL: https://github.com/desh2608/css
• Owner: desh2608
• License: mit
• Created: 2021-11-10T18:49:51.000Z (over 3 years ago)
• Default Branch: main
• Last Pushed: 2022-10-05T14:56:32.000Z (over 2 years ago)
• Last Synced: 2025-03-25T07:11:20.926Z (2 months ago)
• Language: Python
• Size: 232 KB
• Stars: 13
• Watchers: 0
• Forks: 4
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# css

PyTorch implementation of Continuous Speech Separation

>>> This repository is under development.

This repository contains training and inference pipelines for continuous speech

separation on long recordings, similar to that in the [LibriCSS paper](https://arxiv.org/abs/2001.11482).

Salient features:

* On-the-fly training data creation using [Lhotse](https://github.com/lhotse-speech/lhotse).

* Conformer and BLSTM encoders.

* Multi-node training is based on Matthew Wiesner's [nnet_pytorch](https://github.com/m-wiesner/nnet_pytorch). It is a naive form of multi-node training

suitable for clusters where DDP may be slowed down due to an I/O bottleneck.