A Comparison of Three Language Models for Speaker-Dependent Chinese Speech Recognition

Abstract

In this paper, we present a speech recognition system consisting of a signal processing component and a language model. The signal processing component uses traditional techniques such as linear precedence coefficient (LPC), Cepstrum, VQ and HMM. The three language models used in this study are those of context-free grammar, bigram and HMM. The target speeches to be recognized are Chinese Logo (CLogo) program codes. Since CLogo is a formal programming language, we can write down its context-free grammar rules by hand. For bigram model and HMM, their probabilistic linguistic knowledge are automatically learned from a corpus of CLogo programs. All three language models can output the best character combination for the spoken sentence. Empirical results show that CFG performs better than others, with a weakness is that it must be constructed by hand. On the other hand, HMM and bigram can be trained automatically with a corpus. The system is also tested with a corpus of natural language texts of elementary geometry problems.