语音识别中基于神经网络的矢量量化方法

本文对神经网络语音识别中的语音特征提取、网络结构以及学习算法进行了初步的研究，提出了一种用于时特征矢量量化的简化和改进的自组织神经网络模型ＶＱＮＮ。ＶＱＮＮ中引入了动态规划法估计语音样本矢量的码本类中心初值并确定网络的初始权矩阵，可构造出２５６个量化等级的码本矢量。该方法具有较强的鲁棒性且矢量量化过程简单迅速。对２８个地名的语音量化识别实验结果表明了这种量化方法对时识别的有性。

A VECTOR QUANTIZATION APPROACH BASED ON THE NEURAL NETWORK IN SPEECH RECOGNITION

This paper investigates the speech feature extraction , the neural network structure and the learning algorithm for the neural network speech recognition, proposes a simplifing and improving self organization neural network model that is used to speech feature vector quantization VQNN. In VQNN, the dynamic programming method is introduced to estimate the initial value of the speech sample vector cluster center and to determine initial weight matrix, and the code vector that is 256 quantization degrees is built. This method possesses more robust and the vector quantization process is simple and quick. The experiment results for the speech quantization recognition of 28 place name demonstrate the efficiency of this quantization method for speech recognition.