矢量量化降低BP网规模的方法

在用ＢＰ网进行语音和说话人识别过程中，ＢＰ网的输入节点数一般在几百个左右，使得网络的规模过大，训练速度过慢，为此本文在对语音特征进行了有效编码的基础上，充分考虑到ＢＰ网输入的自适应性多维码字间距离与一维码号间距离的不一致性，对量化码间中距离的不一致性，对量码字的码号进行有效的码号变换，用变换后的码号数据经归一化后作为ＢＰ网的输入，从而大大压缩了网络的规模。所进行的语音识别实验及与其它语音识别方法的

Decreasing BP Network Scale Method Based on Vector Quantization

In the process of speech and speaker recognition with BP neural network, nearly several hundreds of input nodes are required, that lead a very large network scale and a very slow train speed. This paper,doing vector quantization and encoding to speech and speaker features, suggests a very effective codenumber transformation method on a deep consideration of the adaptation of BP network to network inputs and the unconsistency of the distance beteen p dimentional codewords and the distance between one dimentional codenumbers. Letting the transformed codenumber as the inputs of BP network, the network scale is compressed in a very large scale. Some speech recognition experiments have been done. The comparisons to the other speech recognition methods result in a very high promotion in train speed, memory capacity needed,and recognition rate.