Speech Recognizer Adaptation: Recognizer Adaptation by Acoustic Model Interpolation

This book focuses on the adaptation of speech recognizers to noisy or reverberant environment. Therefore, three corpora in different noise and reverberation levels are presented. Speech recognition is used. Basics are omitted. As features Mel Frequency Cepstrum Coefficients (MFCC) and several variants of the TempoRAl Patterns (TRAPs) are employed. In order to improve speech recognition even further the following speech recognizer adaptation techniques are Methods like maximum a posteriori (MAP), maximum likelihood linear regression (MLLR), and constrained MLLR (CMLLR) are described in detail. Moreover, the Baum-Welch algorithm is used to interpolate the transition probabilities of the hidden Markov models of the speech recognizer. By application of the adaptation techniques and artificially reverberated data a significant improvement of the recognition rate is achieved.