Influence of Different Speech Representations and HMM Training Strategies on ASR Performance

Abstract

This work studies the influence of various speech signal representations and speaking styles on the performance of automatic speech recognition (ASR).  The efficiency of two approaches to hidden Markov model (HMM) training are compared.Common MFCC and PLP features were exposed to two sources of disturbance applied to the original wide-band speech: (i) stress (Lombard effect) and (ii) transfer channel distortion (simulated telephone line). Subsequently, the efficiencies of the two training strategies were evaluated. Finally, a study of the optimal number of training iterations is introduced.