多权值神经元网络仿生模式识别方法在低训练样本数量非特定人语音识别中与HMM及DTW的比较研究  

Comparison of Biomimetic Pattern Recognition,HMM and DTW for Speaker-Independent Speech Recognition

作　　者：覃鸿[1] 王守觉[1] 

机构地区：[1]中国科学院半导体研究所神经网络实验室,北京100083

出　　处：《电子学报》2005年第5期957-960,共4页Acta Electronica Sinica

基　　金：国家自然科学基金(No.60135010)

摘　　要：　本文将基于多权值神经元网络的仿生模式识别方法用于连续语音有限词汇量固定词组识别的研究中,并将其识别效果与HMM方法及DTW方法进行了比较分析.以15个词组的词汇表做测试,通过调整这三种识别算法的参数,在它们的拒识率相同的情况下,针对参加训练的词汇,比较他们的错误识别率(某类误认为他类);针对未参加训练的词汇,比较他们的错误接受率(误认为某类).结果表明,在低训练样本数量的情况下,仿生模式识别方法能获得更好的识别效果.The purpose of this paper is to compare the performance of three speech recognition methods, one based on Biomimetic Pattern Recognition (BPR) and the other two based on Hidden Markov Models (HMMs) and Dynamic Time Warping (DTW) respectively. As a general purpose model of pattern Recognition, BPR is realized by Multi-Weights Neuron Networks. For the 15 words vocabulary, we analyze the false recognition rate (ratio of accepting a trained word to another trained word) and false acceptance rate (ratio of accepting an untrained word to a trained word) respectively. Experiment results show that when the training data was not sufficient, the manner of BPR achieved a higher performance.

关 键 词：仿生模式识别 多权值矢量神经元 语音识别 HMMs DTW 

分 类 号：TN912[电子电信—通信与信息系统]