Fast on‐line speaker/environment adaptation using modified maximum likelihood stochastic matching

The problem of speaker/environment adaptation to improve the recognition accuracy and thus making recognizers robust is addressed here. For this, a fast on‐line adaptation algorithm that does not need a separate adaptation training data and that adapts acoustic models fast enough to achieve near real‐time recognition is developed. This technique is based on stochastic matching in the model space similar to [A. Shankar and C.‐H. Lee, IEEE Trans. Signal Process. 4, 190–202 (1996)]. For fast adaptation only the models and the mixture components that need to be adapted are selected based on the cluster formation and Euclidean distance. This adaptation algorithm is implemented as part of a GMM based continuous speech recognizer. It is tested using a non‐native speakers dataset. For example, the five best hypotheses output of the speech recognizer before and after applying the adaptation technique indicated that the right answer corresponding to an utterance ‘‘none of the earth’’ before adaptation did not correspond to the best hypothesis; however, it corresponded to the third best. After adaptation all of the five‐best hypotheses converged to the right answer. The test results of this technique on a larger non‐native speakers’ dataset shows 70% to 75% relative WER improvement.