A Comprehensive Study on CTC Loss for Timing Alignment

This repo contains the essential code for the ICASSP 2022 submission: "A Comprehensive Study on CTC Loss for Timing Alignment", Xingyu Cai, Jiahong Yuan, Renjie Zheng, Kenneth Church.

Credits

The code is largely based on the following repositories and corresponding references:

• Phoneme recognition: https://github.com/huggingface/transformers/tree/master/examples/research_projects/wav2vec2
• Python CTC: https://github.com/vadimkantorov/ctc
• Entropy regularized CTC [1]: https://github.com/liuhu-bigeye/enctc.crnn

Install required packages

pip install requirements.txt

Minimum effort to run

bash script.sh standard no regular

The script runs for CTC using a combination of standard update scheme, no prior and regular vocabulary It runs on the TIMIT dataset and obtain phoneme-error-rate (PER) and alignment error (MSE and MAE).

Reference

1. Connectionist Temporal Classification with Maximum Entropy Regularization Hu Liu, Sheng Jin and Changshui Zhang. Neural Information Processing Systems (NeurIPS), 2018.