有限词汇、特定人、商品化的英语语音识别系统

A Limited Vocabulary Speaker-Dependent Marketable English Speech Recognition System

介绍了一个在微机上实现的有限词、特定人语音识别系统．该系统采用连续Ｍ元高斯混合密度的隐式马尔柯夫模型（ＣＤＨＭＭ）为识别方法，以修改后的Ｂａｕｍｗｅｌｃｈ方法为训练重估算法．文中提出了对语音特征矢量非线性归一化预处理，和对训练数据不足的ＨＭＭ模型特征空间进行后处理修正的算法，还提出一种基于语音知识的模型初始化方法．经实验证明，系统的识别率可以达到９０％以上．

This paper introduced a limited vocabulary speaker-dependent marketable English speech recognition system running on a microcomputer-386 using VITERBI algorithm in recognition process under the Continuous Density Hidden Markov Model (CDHMM). Particularly, we propose a modified Baum-Welch algorithm as training algorithm, a nonlinear programming method to proceed time warping about feature vector, and a modification to the parameter of CDHMM for resolving insufficient training data. Besides, we have presented a method based on speech knowledge to initialize model. The experiments show that the recognition rate of our system is more than 90%.