Skip to content

Alexander-H-Liu/NPC

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Non-Autoregressive Predictive Coding

This repository contains the implementation of Non-Autoregressive Predictive Coding (NPC) as described in the preprint paper submitted to ICASSP 2021.

A quick example for training NPC

python main.py --config config/self_supervised/npc_example.yml \
               --task self-learning

Some notes

  • We found the unmasked feature produced by the last ConvBlock layer a better representation. In the phone classification tasks, switching to the unmasked feature (PER 25.6%) provided a 1.6% improvement over the masked feature (PER 27.2%). Currently, this is not included in the preprint version and will be updated to the paper in the future. Please refer to downstream examples to activate this option.

  • APC/VQ-APC are implemented with the following modifications for improvement (for the unmodified version, please visit the official implementation of APC / VQAPC)

    • Multi-group VQ available for VQ-APC, but with VQ on last layer only

    • Using utterance-wised CMVN surface feature(just as NPC did)

    • Using Gumbel Softmax from official API of pytorch

  • See package requirement for toolkits used, tensorboard can be used to access log files in --logdir.

Contact

Feel free to contact me for questions or feedbacks, my email can be found in the paper or my personal page.

Citation

If you find our work and/or this repository helpful, please do consider citing us

@article{liu2020nonautoregressive,
  title   = {Non-Autoregressive Predictive Coding for Learning Speech Representations from Local Dependencies},
  author  = {Liu, Alexander and Chung, Yu-An and Glass, James},
  journal = {arXiv preprint arXiv:2011.00406},
  year    = {2020}
}

About

Non-Autoregressive Predictive Coding

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published