The importance of prosodic factors in phoneme modeling with applications to speech recognition

Abstract

This paper tests speech recognition using prosody dependent allophone models. The log likehoods of various prosodically labeled phonemes are calculated using Baum-Welsh re-estimation. These log likehoods are then compared to log likehoods of non-prosodically labeled phonemes. Based on the comparison of these log likehoods, it can be concluded that modeling all prosodic information directly in the vowel model leads to improvement in the model. Consonants, on the other hand, split naturally into three categories, strengthened, lengthened and neutral.