基于高斯相似度分析的插值自适应算法

Interpolation Adaptation Algorithm Based on Gaussian Similarity Analysis

快速说话人自适应算法在非特定人连续语音识别的应用中有重要意义 .现在流行的自适应算法多数只考虑均值的自适应 .本文提出的自适应算法可以快速的对协方差矩阵进行自适应 .该算法是用高斯相似度度量协方差矩阵间的距离 ,并由此测度建立了反映协方差矩阵结构关系的二叉决策树 .树的每个中间节点包含一个类质心 .在决策树基础上 ,训练多个与特定人模型相关的类质心 .自适应时 ,通过对这些类质心进行线性插值得到自适应的协方差矩阵 .实验结果表明 ,该方法能够在仅有一句自适应数据的情况下 ,使系统误识率由 2 9 4 9%下降到 2 7 5 5 % .

Fast speaker adaptation is very important in application of speaker independent continuous speech recognition.Recently speaker adaptation methods almost just update mean vector.A new adaptation algorithm proposed in this paper can adapt covariance matrixes rapidly.Gaussian similarity is used for measuring the distance of different covariance.A binary decision tree is constructed with this measure.And each middle node includes a covariance matrix of cluster center.Lots of covariance matrixes corresponding to speaker dependent model are trained based on this tree.During adaptation,covariance matrixes are updated through linear interpolating those covariances of cluster center.It can be shown from the experiments that error rate is descended from 29 49% to 27 55% in case of just one adapted sentence.