Discriminant Training of Front-End and Acoustic Modeling Stages to Heterogeneous Acoustic Environmen

Automatic Speech Recognition (ASR) still poses a problem to researchers. In particular, most ASR systems have not been able to fully handle adverse acoustic environments. Although a large number of modifications have resulted in increased levels of performance robustness, ASR systems still fall short of human recognition ability in a large number of environments. A possible shortcoming of the typical ASR system is the reliance on a single stream of front-end acoustic features and acoustic modeling feature probabilities. A single front-end feature extraction algorithm may not be capable of maintaining robustness to arbitrary acoustic environments. Acoustic modeling will also degrade due to distributional changes caused by the acoustic environment. This thesis explores the parallel use of multiple front-end and acoustic modeling elements to improve upon this shortcoming. Each ASR acoustic modeling component is trained to estimate class posterior probabilities in a particular acoustic environment. In addition to discriminative training of the probability estimator, existing feature extraction algorithms are modified in such a way as to improve class discrimination in the training environment. More specifically, Linear Discriminant Analysis provides a mechanism for obtaining discriminant temporal basis functions that can replace components of the existing algorithms that were designed in either an empirical or intuitive manner. Probability streams are generated using multiple front-end acoustic modeling stages trained to heterogeneous acoustic environments. In new sample acoustic environments, simple combinations of these probability streams give rise to word recognition rates that are superior to the individual streams.