| Presentation # | 12 |
| Session: | ASR IV |
| Location: | Kallirhoe Hall |
| Session Time: | Friday, December 21, 13:30 - 15:30 |
| Presentation Time: | Friday, December 21, 13:30 - 15:30 |
| Presentation: |
Poster
|
| Topic: |
Speech recognition and synthesis: |
| Paper Title: |
Multilingual sequence-to-sequence speech recognition: Architecture, transfer learning, and language modeling |
| Authors: |
Jaejin Cho, Johns Hopkins University, United States; Murali Karthick Baskar, Brno university of technology, Czech Republic; Ruizhi Li, Matthew Wiesner, Johns Hopkins University, United States; Sri Harish Mallidi, Amazon, United States; Nelson Yalta, Waseda University, Japan; Martin Karafiat, Brno university of technology, Czech Republic; Shinji Watanabe, Johns Hopkins University, United States; Takaaki Hori, Mitsubishi Electric Research Laboratories, United States |
| Abstract: |
Sequence-to-sequence (seq2seq) approach for low-resource ASR is a relatively new direction in speech research. The approach benefits by performing model training without using lexicon and alignments. However, this poses a new problem of requiring more data compared to conventional DNN-HMM systems. In this work, we attempt to use data from 10 BABEL languages to build a multilingual seq2seq model as an initial model, and then perform several transfer learning approaches across 4 other BABEL languages. We also explore different architectures for a multilingual seq2seq model to improve their performance. Further analysis is performed to understand the importance of scheduled sampling approach to bring the model distribution closer to the target data distribution. The paper also discusses about the effect of combining a recurrent neural network language model (RNNLM) with a seq2seq model during decoding. Experimental results show that the multilingual transfer learning model shows substantial gains over monolingual models across all 4 BABEL languages. Incorporating an RNNLM also brings significant gains, and achieves recognition performance comparable to the models trained with twice more training data. |