摘要
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 proposes a novel method named sequential cluster network to solve the time alignment problems in artificial neural network (ANN) based speech recognition. Using this network, a fixed number of feature vectors is extracted from the input speech signal, and then processed by a ANN classifier for recognition. Compared with other ANN based methods, the proposed method has many advantages such as less time taken for training and recognition with simpler ANN structure and higher accuracy. A word recognition system is established based on this method and then tested with two sets of English words. The experiment results demonstrate that the proposed method outperforms the conventional HMM and some other ANN methods.
出处
《电路与系统学报》
CSCD
2000年第2期99-103,共5页
Journal of Circuits and Systems
关键词
神经网络
语音识别
有序聚类
Time Wrapping Algorithm , Neural Network , Speech RecognitionH
