Improvement in text-dependent mispronunciation detection for English learners

Abstract

This paper put forth two novel approaches to effectively improve the performance of mispronunciations detection in English learners speech. On one hand, a distance measure called Kullback–Leibler Divergence (KLD) between Hidden Markov Models (HMMs) is introduced to optimize the probability space of a posteriori probability; On the other hand, back end processing of normalization based on the variants of speakers is introduced to improve the performance of the system. Experiments on a database of 6360 syllables pronounced by 50 speakers with varied pronunciation proficiency indicate the promising effects of these methods by decreasing the FRR from 58 to 44% at 20% FAR.