Tonal phoneme based model for Vietnamese LVCSR

Abstract

This paper proposes an algorithm that is first known as a grapheme-to-phoneme method to transform any Vietnamese word to a tonal phoneme-based pronunciation. The tonal phoneme set produced by this algorithm is further used to develop some acoustic models which integrated tone information and tonal feature. The processes using the Kaldi toolkit to develop a LVCSR system and extract a bottleneck feature which is calculated from a trained deep neural network for Vietnamese are also presented. The results showed that the use of tonal phoneme improved by 1.54% of word error rate (WER) compared to the system using the nontonal phoneme, the use of tonal feature information improved by 4.65% of WER, and of the bottleneck feature gave the best WER with about 10% improvement.