Institute of Information Science Academia Sinica
Topic: High-rank LSTM-CTC Acoustic Modeling for ASR and RNNLM with Trust Regularization
Speaker: Dr. Mei-Yuh Hwang (Mobvoi AI Lab)
Date: 2019-04-18 (Thu) 10:30 – 12:30
Location: Auditorium106 at IIS new Building
Host: Hsin-Min Wang

Abstract:

This is a preview of our two ICASSP 2019 papers. The 1st paper talks about how we use mixture of experts in LSTM-CTC acoustic modeling for automatic speech recognition (ASR). The 2nd paper talks about how knowledge distillation with trust regularization improves perplexity of RNNLM, and hence lowers ASR error rates.


BIO:

Mei-Yuh received her PhD in Computer Science from Carnegie Mellon University in December 1993, in acoustic modeling for speech recognition. She worked for Microsoft Research and Product divisions, and University of Washington after graduation until 2016. She is serving as the director of Mobvoi AI Lab in Seattle, Washington. She has published numerous conference and journal papers, and industry patents in the field of speech recognition, machine translation, and natural language understanding and dialog management. She is elected as an IEEE Fellow, class of 2019, for the contributions to speech and language technology.