Improvement of phone recognition accuracy using source and system features

Abstract

The goal of this work is to improve phone recognition accuracy using combination of source and system features. As speech is produced by exciting time varying vocal tract system with time varying excitation, we want to explore both source and system components of speech production system for phone recognition. The excitation source information is derived by processing linear prediction residual of speech signal. Melfrequency cepstral coefficient features are used for capturing vocal tract information. The Phone Recognition Systems (PRSs) are developed using hidden markov models. The proposed PRSs are developed for English and an Indian language Bengali using TIMIT and Phonetic, Prosodically Rich Transcribed speech corpora, respectively. We have also developed tandem PRSs using the phone posteriors obtained from feedforward neural networks. The tandem PRSs developed using combination of excitation source and system features, outperform the conventional tandem systems developed using system features alone.