一种CDHMM／MLP混合模型及其在语音识别中的应用

本文提出了一种由连续隐马尔可夫模型与多层感知器构成的混合模型，并将该模型应用于语音孤立词识别，这种混合模型首先用ＣＤＨＭＭ来获取输入信号的动态特性，然后再以ＭＬＰ分类器对输入信号进行分类识别。其主要目的是通过ＭＬＰ分类器，对ＣＤＨＭＭ中的似然估计值进行分析，分类，以加强和提高ＣＤＨＭＭ的分类能力。根据这种混合模型，我们建立了一个含３０个英语单词的语音识别系统。实验结果表明，该系统的识别率明显高于传

An CDHMM/MLP Hybrid Model and Its Application to Speech Recognition

This paper presents a hybrid model of Continuous Density Hidden Markov Model (CDHMM) and the Multi-Layer Perceptron (MLP). According to this model, CDHMM is used to capture the dynamics of input speech signal and then the captured feature vectors are discriminated by the MLP classifier. The main purpose of this model is to improve the recognition rate of CDHMM by means of the strong performance of MLP to classify the likelihood values withdrawn from CDHMM. Recognition experiment with 30 isolated English words demonstrated the advantage of proposed model over the conventional CDHMM.