用矢量量化和隐马尔可夫模型实现英语话句的识别

描述用矢量量化和隐马尔可夫模型实现的英语话句识别系统.采用逐级优化分裂聚类分析获取矢量量化的码本,用平均振幅函数及过零率进行单词切分,用线性预测参数的似然比距离衡量两个矢量差异的大小,使用一阶从左至右的隐马尔可夫模型,用多个输出符号序列进行训练,用Viterbi算法进行识别.用文法分析技术辅助实行识别结果的判定.

RECOGNITION OF SPOKEN ENGLISH SENTENCES WITH VECTOR QUANTIZATION AND HIDDEN MARKOV MODELS

This paper describes a system which recognizes spoken english sentences with Vector Quantization and Hidden Markov Models. The Codebook of Vector Quantization is obtained using an optimized binary-split cluster algorithm. Words in sentences are segmented by zero-crossing ratio and average magnitude function. The differences between vectors are measured in terms of likelihood-ratio-distance of LPC parameter. One-order left-to-right Hidden Markov Models are used. The models are trained with a set of obscrvation sequences spoken words are recognized using Viterbi algorithm. With several efficient techniques a fairly fast recognition system is implemented on microcomputer in high-level language. In the decision phase, some grammatical constraints are used to the input sentences, resulting in a great improvement to recognition accuracy.