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.
Salient features:
- On-the-fly training data creation using Lhotse.
- Conformer and BLSTM encoders.
- Multi-node training is based on Matthew Wiesner's 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.