关键词: 隐马尔可夫模型, 语音识别, 高效简约系统, 声学模型拓扑结构, 贝叶斯信息准则, 粒群优化

Abstract: Aiming at the current situation of lacking standard methods to construct efficient and concise speech recognition systems, a new method to build this kind of systems is proposed, which involves determining the relative importance of the recognition performance and system complexity through the value of the regularization coefficient in Bayesian Information Criterion（BIC） and optimizing acoustic model topologies using improved Particle Swarm Optimization（PSO） algorithm. Experiments on TIDigits corpus show that the new system constructed by using this method obtains 7.85% absolute increase in sentence correct rate compared to the baseline with the same complexity and built in the conventional way, and reduces 51.4% of system complexity compared to the baseline with the same recognition rate, which indicates that the new system is capable of obtaining higher recognition rate with lower system complexity.

Key words: Hidden Markov Model（HMM）, speech recognition, efficient and concise system, acoustic model topology, Bayesian Information Criterion（BIC）, Particle Swarm Optimization（PSO）