Acoustic HMM Compensation for ASR on Mobile Devices

In the talk I gave a summary of the work performed during my stay at AT&T Labs-Research. The aim of the work was to investigate methods that have potential of improving robustness of speech recognition on mobile communications devices. The work was performed using a speech database that was collected from a directory retrieval task using a simulation of a hand-held mobile device. The data was recorded using both close-talking and device-mounted microphones in office and cafeteria acoustic environments. Since it is generally assumed that a mobile communications device is dedicated to a particular user, the techniques for "personalizing" a speaker independent automatic speech recognizer to a given user and a given device were investigated. These techniques include frequency warping based speaker normalization and speaker adaptation using maximum likelihood linear regression. Mobility also implies exposure to harsh and unpredictable acoustics environments. The HMM compensation techniques were investigated to deal with both slowly varying and rapidly varying components of acoustic background noise. Parallel model combination (PMC), a technique for integrating models of the acoustic background environment with speech model, was applied in conjunction with regression based channel compensation procedures. Jacobian matrix based compensation was used to describe how perturbational changes in noise characteristics result in perturbational changes to speech model parameters. This technique was applied to compensate for the effects of rapidly varying background conditions.