Abstract
Most successful speech recognition systems are based on Hidden Markov Models (HMM), which rely on computation-intensive Viterbi search during recognition. The first part of this thesis focuses on the optimization of Vierbi beam search in HMM decoding for isolated-word speech recognition. The proposed data-driven method can effectively identify a near-optimal beam search ranking curve that can reduce the computation time to an acceptable amount while minimizing the reduction in recognition rate based on a set of sample data. Experimental results based on the most famous 300 poems in Tang Dynasty of China demonstrate the feasibility of the proposed approach. In the second part of this thesis, we applied the proposed approach to a multilingual speech recognition system that can deal with Mandarin Chinese, Taiwanese, English, and their combinations. The system employs different acoustic models for different languages, and hence possesses a high degree of flexibility and modularity. We also described how to treat similar phone models as an equivalence class in order to reduce the total number of phone models for a limited speech corpus. Experimental results demonstrated its feasibility for automatic speech recognition for fixed-domain vocabularies.