改进的MKM算法及其在语音识别中的应用

A Modified MKM Clustering Algorithm and Its Application in Speech Recognition

提出了一种用于矢量量化的改进的聚类算法,该算法在MKM(Modified K-Means)算法的框架的基础上,对初始码本的生成、失真测度的选择、非典型胞腔的处理等方面进行了改进,从而减少了原算法在能量和增益上对聚类结果的影响。并将该算法应用于波形编辑孤立字识别器,这种识别器直接对语音样本的时域波形进行训练和聚类,不需要提取语音参数,算法复杂度较低,加上提出的聚类算法失真测度简单易实现,对芯片的运算能力要求不高,非常适用于有低成本要求的语音识别器场合。通过中文元音字识别的实验证明,在相同码本尺寸下,运用改进后的聚类算法的识别器的识别率有所提高。

A modified clustering algorithm used in vector quantization is presented in this paper.It is based on the framework of MKM clustering algorithm,and modifies some aspects in order to lower down the effect of power and gain,e.g.initial codehook obtainment,distortion measure,the method of dealing with unrepresentative Voronoi Cells,etc. Using the proposed VQ algorithm,we have developed a simple isolated word speech recognizer.It simply trains and clusters the waveform samples of the speech,which does not request high calculation capability and suitable for the low cost application.Experiments on Chinese vowel recognition show that with the modified clustering algorithm the recognition accuracy of the speech recognizer is higher.