MPhil Thesis Proposal Defence


"Development of an Asynchronous Multi-band System for Continuous Speech
Recognition"

By

Mr. Yik-Cheung Tam


Abstract

Recently, multi-band automatic speech recognition (MBASR) has been
proposed by Bourlard et al. and Hermansky et al. to improve robustness
under noisy environment. It is motivated by the empirical findings by
Harvey Fletcher of Bell Labs from a thorough study of human speech
recognition(HSR) in which partial speech recognition in sub-bands is
believed to take place and then the sub-band decisions are recombined to
arrive at a global decision. They found that the full band error rate is
empirically equal to the multiplication of sub-band error rates. This
implies that human can recognize the speech correctly if there exists a
correct sub-band recognition.

The MBASR framework proposed by Bourlard et al. and Hermansky et al. is to
divide the full frequency band into sub-bands and a speech recognizer is
built for each sub-band. During recognition, decisions from individual
sub-band recognizers are recombined to arrive at a final decision at some
phonetic or linguistic level.

In this thesis, a multi-band system is implemented with several proposed
features. First, it can be used for continuous speech recognition which is
generally a basic requirement for research or real-life application
deployment. Second, it allows asynchronous recombination of sub-band
information at any desired units such as phoneme/syllable/word units.
Third, sub-band information are recombined in an optimal sense.

In this proposal, HMM composition framework is introduced as the back-bone
in a multi-band system to address the sub-band asynchrony issue. In
addition, continuous speech recognition can easily be realized under such
framework. Using  linear recombination of sub-band log-likelihoods,
string-based minimum classification error criterion (MCE) is employed to
optimize the sub-band weightings using simulated noisy speech. Our first
preliminary experiment suggested that word-based MCE weight training is
effective in emphasizing the more reliable sub-band with more weights
resulting in boosted performance on an isolated digit recognition task.
Our second preliminary experiment suggested that the heuristics of
assuming uniform transition probabilities in composite HMMs results in
performane degradation on a connected digit recognition task when
asynchrony is allowed. It is crucial to either compute them using sub-band
counterparts or re-train them in order to utilize the HMM composition
framework successfully.


Date:				Thursday, 28 December 2000

Time:				3:00p.m.-5:00p.m.
	
Venue:				Room 2302
				Lifts 17-18

Chairman:			Dr. Cunsheng Ding

Committee Members:		Dr. Brian Mak (Supervisor)
				Dr. James Kwok
				Dr. Man-Hung Siu


**** ALL are Welcome ****