The Johns Hopkins University

Whiting School of Engineering

Department of Electrical and Computer Engineering

Multiple Heteroscedastic Linear Discriminant Analysis

Seminar By

Haolang Zhou

Graduate Research Assistant

Electrical and Computer

Abstract:

Current automatic speech recognition systems employ dimension reduction schemes to project high dimensional features to a lower dimensional space. However, the resulting lower dimensional features used during recognition and decoding are the same for all phones. At the same time, for related tasks such as phone recognition, phoneticians have long relied on phone-specific cues to verify the identity of different phones.

In this proposal, we present a novel feature reduction mechanism that allows different transforms for each phone, class of phones, or even acoustic state. The approach adopted is to provide a rich and redundant representation of the input feature space which is then reduced using class-specific linear transforms. Care needs to be taken to ensure that the resulting scores are comparable across different phones. Performance is reported on Rich Transcription 04 Arabic speech recognition test data. Preliminary results are promising.

Friday, October 12, 2007

11:00 AM

Barton 114

Refreshments will be served at 10:45 AM

FOR DISABILITY INFORMATION

CONTACT: Candace Abel (410) 516-7031 cabel@jhu.edu