论文部分内容阅读
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，提出了一种用于语音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语音识别的有效性．
In this paper, speech feature extraction, network structure and learning algorithm in neural network speech recognition are studied preliminarily. A simplified and improved self-organizing neural network model VQNN for speech feature vector quantization is proposed. In VQNN, a dynamic programming method is introduced to estimate the initial value of codebook center of speech sample vectors and to determine the initial weight matrix of the network, so that 256 quantization codebook vectors can be constructed. The method is robust and the vector quantization is simple and fast. The experimental results of quantifying speech quantification of 28 place names demonstrate the effectiveness of this method for speech recognition.