Multiple acoustic cues for Korean stops and automatic speech recognition

Abstract

The purpose of this thesis is to analyse acoustic characteristics of Korean stops by way of multivariate statistical tests, and to apply the results of the analysis in Automatic Speech Recognition (ASR) of Korean. Three acoustic cues that differentiate three types of Ko¬ rean oral stops are closure duration, Voice Onset Time (VOT) and fundamental frequency (FO) of a vowel after a stop. We review the characteristics of these parameters previously reported in various phonetic studies and test their usefulness for differentiating the three types of stops on two databases, one with controlled contexts, as in other phonetic stud¬ ies, and the other a continuous speech database designed for ASR. Statistical tests on both databases confirm that the three types of stops can be differentiated by the three acoustic parameters. In order to exploit these parameters for ASR, a context dependent Hidden Markov Model (HMM) based baseline system with a short pause model is built, which results in great improvement of performance compared to other systems. For mod¬ elling of the three acoustic parameters, an automatic segmentation technique for closure and VOT is developed. Samples of each acoustic parameter are modelled with univariate and multivariate probability distribution functions. Stop probability from these models is integrated by a post-processing technique. Our results show that integration of stop prob¬ ability does not make much improvement over the results of a baseline system. However, the results suggest that stop probabilities will be useful in determining the correct hy¬ pothesis with a larger lexicon containing more minimal pairs of words that differ by the identity of just one stop.