TR2019-006

Promising Accurate Prefix Boosting for Sequence-to-Sequence ASR


    •  Baskar, M.K., Burget, L., Watanabe, S., Karafiat, M., Hori, T., Cernocky, J.H., "Promising Accurate Prefix Boosting for Sequence-to-Sequence ASR", IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), DOI: 10.1109/​ICASSP.2019.8682782, May 2019, pp. 5646-5650.
      BibTeX TR2019-006 PDF
      • @inproceedings{Baskar2019may,
      • author = {Baskar, Murali Karthick and Burget, Lukas and Watanabe, Shinji and Karafiat, Martin and Hori, Takaaki and Cernocky, Jan, Honza},
      • title = {Promising Accurate Prefix Boosting for Sequence-to-Sequence ASR},
      • booktitle = {IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP)},
      • year = 2019,
      • pages = {5646--5650},
      • month = may,
      • doi = {10.1109/ICASSP.2019.8682782},
      • issn = {2379-190X},
      • isbn = {978-1-4799-8131-1},
      • url = {https://www.merl.com/publications/TR2019-006}
      • }
  • Research Areas:

    Artificial Intelligence, Machine Learning, Speech & Audio

Abstract:

In this paper, we present promising accurate prefix boosting (PAPB), a discriminative training technique for attention based sequence-tosequence (seq2seq) ASR. PAPB is devised to unify the training and testing scheme effectively. The training procedure involves maximizing the score of each partial correct sequence obtained duringbeam search compared to other hypotheses. The training objective also includes minimization of token (character) error rate. PAPB shows its efficacy by achieving 10.8% and 3.8% WER with and without external RNNLM respectively on Wall Street Journal dataset.

 

  • Related News & Events