One project focused extended previous work in maximum likelihood channel estimation by introducing a prior distribution model of the channel/noise and using Bayesian estimation techniques and by assessing methods with training data compensation. Both cepstral filtering and model adaptation techniques are explored to assess computation-performance tradeoffs. A progressive recognition search allows for phone-conditioned channel estimation to improve compensation for short utterances (likely in many applications). Up to 10% error rate reduction is observed. The approach can easily be implemented in any multi-pass speech recognition system.
A second project involves development of speaker separation techniques explicitly aimed at improving automatic speech recognition, with a secondary goal of using recognition to improve the separation algorithm for improved perceptual quality.
(March 1994 - July 00)
SPONSORS: National Science Foundation and ARPA, NSF IRI-9408896