CTC LOSS FUNCTION WITH A UNITLEVEL AMBIGUITY PENALTY

CTC LOSS FUNCTION WITH A UNIT-LEVEL AMBIGUITY PENALTY Ryoichi Takashima, Sheng Li and Hisashi Kawai National Institute of Information and Communications Technology, Japan ICASSP 2018

Outline Introduction CTC loss with ambiguity penalty Experiments Conclusion

Introduction Language model P(W) Language model arg max(P(X|W)P(W)) Decoding Acoustic model can also not use the language model Hello Decoding Acoustic model P(X|W) HMM S 1 S 2 Sn DT SAT DNN CD Senone GMM(frame level alignment) Features(x 1) Features(x 2) Features(xn) CI Phone CTC learn acoustic frames to character mappings DBRNN/LSTM Decision Tree Features(x 1) Dict Features(x 2) Features(xn) + Ambiguity penalty

CTC loss with ambiguity penalty CTC AP

CTC loss with ambiguity penalty CTC AP

CTC loss with ambiguity penalty

Experiments CTC model Toolkit: EESEN Feature: Fbank Optimization algorithm: SGD with momentum 0. 9 Corpus: 1. WSJ 0(15 hr train_si 84) 2. WSJ 0+WSJ 1(81 hr train_si 284) 3. CHi. ME 4 Lexicon: • WSJ CMU dictionary and the 20, 000 -word vocabulary WSJ pruned language model(lm_tgpr) • CHi. ME 4 CMU dictionary and the 50, 000 -word vocabulary WSJ pruned language model(lm_tgpr_5 k)

Experiments

Experiments

Conclusion In experiments using multiple corpora, our proposed modification did not outperform conventional CE-DNN-HMMs, but did improve the WERs of conventional CTC-based models when there were few training data. This means that the proposed method did not realize the full potential of CTC-based models, but did improve their data-efficiency. In future work, we will evaluate our pro- posed method in these tasks, and continue efforts to achieve the full potential of CTC-based models.