论文部分内容阅读
本文提出了一种新的网络结构,我们称之为有序聚类网络。这种网络能够对语音信号进行特征提取,很好地解决神经网络语音识别中的时间规整问题。有序聚类网络从输入语音信号的特征矢量序列中提取出一组固定数目的特征矢量,然后将这组特征矢量馈入神经网络分类器进行识别。和其他的神经网络语音识别方法相比较,用这种网络进行前端处理,可以缩短后端神经网络分类器的训练和识别时间,简化分类器的网络结构并保持较高的识别率。根据该方法我们建立了一个语音识别系统,并对两组英语单词进行了识别测试。实验结果表明,该方法优于传统的隐马尔可夫模型方法以及其它一些神经网络方法。
This paper presents a new network structure, which we call an ordered clustering network. This kind of network can extract the feature of speech signal and solve the problem of time warping in neural network speech recognition well. An orderly clustering network extracts a fixed number of eigenvectors from the eigenvector sequence of the input speech signal, and then feeds the set of eigenvectors into a neural network classifier for identification. Compared with other neural network speech recognition methods, the front-end processing using this network can shorten the training and recognition time of the back-end neural network classifier, simplify the network structure of the classifier and maintain a high recognition rate. According to this method, we establish a speech recognition system, and test two groups of English words. Experimental results show that this method is superior to the traditional Hidden Markov Model and other neural network methods.