Modeling of non-native speech automatic speech recognition

Abstract

Heavily accented non-native speech represents a significant challenge for automatic speech recognition (ASR). Globalization again emphasizes the urgent of the research to address these challenges. In the thesis, the author will focus on the two core parts of ASR system to improve the accuracy of the non-native speech recognition. Firstly, the author will attempt to improve the acoustic model of ASR. Non-native acoustic modelling is an essential component in many practical ASR systems. The author has attempted MLLR and MAP adaptation, and Neural Network techniques, and the combine of them, and has integrated both the target and source languages information into one model. Secondly, the author also explores the issue in the lexicon model. Statistical method is used to find out the probability of the non-native-speaking pronunciation and form new dictionary.