Boosting Thai Syllable Speech Recognition Using Acoustic Models Combination

Abstract

In this paper, a highly effective system for Thai speech recognition is proposed. The speech recognizer for so-called speaker-independent is created by using Continuous Density Hidden Markov Model (CDHMM). In the acoustic level, the models trained for both speaker genders, and for each separate gender are investigated and tested in terms of accuracy. Experimental evaluation shows that with the acoustic models combination, the accuracy could be improved considerably in the acoustic level. The acoustic combination can support spoken utterances from both genders and still provide the high accuracy simultaneously. Interestingly, when using the acoustic models combination, the syllable accuracy of 89.84% is achieved with 4.53% improvement over using the conventional acoustic models trained for both genders.