Codes for paper:

[Interspeech 2021] End-to-End Language Diarization for Bilingual Code-switching Speech

I am trying to include the data preprocessing steps in this repo so that you can reproduce the results faster. But maybe after I complete my work for interspeech 2022. Hope our work help your research. -- Updated 2022 Jan

Pls cite as follow if you referred to this work:

@inproceedings{liu21d_interspeech,
author={Hexin Liu and Leibny Paola García Perera and Xinyi Zhang and Justin Dauwels and Andy W.H. Khong and Sanjeev Khudanpur and Suzy J. Styles},
title={{End-to-End Language Diarization for Bilingual Code-Switching Speech}},
year=2021,
booktitle={Proc. Interspeech 2021},
pages={1489--1493},
doi={10.21437/Interspeech.2021-82}
}

Requirement:

argparse
torch
tqdm
numpy
scipy

• There is no script for making data (I am trying to include them soon), pls do it yourself for now and revise the code in "data_load.py" accordingly.
• Note that the torch.cuda.deterministics=True conflicts with dilated conv1d which makes the code very slow, so we set it to False in train_xsa.py and thus don't fix the random seed in the training stage. But if you would like to do that, that's also fine, just set it to True.
• No particular initialization is needed. Just run it. And if your have better hyperparams, pls email me so that I can use them hahaha:)