THE USE OF RECURRENT NEURAL NETWORKS FOR AUTOMATIC SPEECH RECOGNITION: A SURVEY

PhD Qualifying Examination


Title: "THE USE OF RECURRENT NEURAL NETWORKS FOR AUTOMATIC SPEECH 
RECOGNITION: A SURVEY"

by

Mr. Hengguan HUANG


Abstract:

Over the past few years, there has been a resurgence of interest in using 
recurrent neural networks for automatic speech recognition. The use of 
recurrent neural networks is a very natural way of acoustic modeling 
because speech is essentially a dynamic process. Some modern recurrent 
network models, such as LSTM and GRU, have demonstrated promising results 
on this task. However, they suffer from some important drawbacks, 
including limited length of history, scalability of multidimensional 
features and uncertainty handling. Many variants of recurrent neural 
networks have been developed in an attempt to address these drawbacks. We 
classify these variants into three categories based on three key drawbacks 
we have identified. Each technique is described and its application to 
acoustic modeling in automatic speech recognition, if any, is discussed. 
In the last part, we further conclude current variants of recurrent neural 
networks and its advantages and disadvantages and discuss possible 
research directions.


Date:			Thursday, 3 August 2017

Time:                  	11:00am - 1:00pm

Venue:                  Room 2611
                         Lifts 31/32

Committee Members:	Dr. Brian Mak (Supervisor)
 			Dr. Raymond Wong (Chairperson)
 			Prof. Fangzhen Lin
 			Dr. Xiaojuan Ma


**** ALL are Welcome ****