» Pronunciation Modeling of Spontaneous Mandarin Speech Using Phonetic Feature Distance and Optimal Gaussian Mixture Sharing

Liu Yi1, Pascale Fung1, William Byrne2, and Umar Ruhi3

1 Dept. EEE, Hong Kong University of Science and Technology, Hong Kong

2 CLSP/ECE, The Johns Hopkins University , Baltimore MD, USA

3 Dept. CS, University of Toronto, Canada

Presented: May 2001.

Pronunciations in spontaneous, conversational speech tend to be much more varied than in carefully read speech. Pronunciation modeling is an efficient way to improve recognition performance. In this paper, we propose incorporating pronunciation variations into acoustic model training. We present our method of incorporating phonetic feature distance into phone variation probabilities. In addition, we present an efficient criterion for choosing the optimal Gaussian mixture components from surface states, and share the selected Gaussian mixture components with canonical states according to variation probabilities. Experiments showed that phonetic feature distance based state level Gaussian mixture sharing improves syllable accuracy by 4.17% absolutely after re-training. Adding optimal mixture component selection, syllable accuracy improved significantly by 4.78% absolutely with respect to the baseline.