基于动词相似度的特征分类方法

A Feature Classification Method Based on Verb Similarity

本文提出了一种新的说话人特征分类方法,基于计算动词相似度理论,建立距离和趋势的评价模型,通过计算特征向量与k-means算法聚类所得的聚类中心的相似度矩阵,将说话人个性特征从MFCC特征域映射到说话人相似度属性空间中,形成新的特征向量集,这样,每个说话人的特征向量将被聚为在距离和变化趋势上最具相似性的k分类。之后,利用GMM模型在属性空间内进行联合概率分析、匹配,建立新的说话人识别系统。本文采用标准TIMIT语音库与NIST语音库在该识别系统中进行一系列实验,结果表明,该基于新的优化特征分类的识别系统,对比传统的说话人识别系统,在等错误率上有很好的提高。

A new strategy of feature classification method for speaker recognition based on computational verb similarity is presented.On the evaluation model with the similarity function of both distance and trend in reel cepstral domain, the new feature vectorsets were assorted after comparing clustering centers, which were obtained by uti]izing k-means algorithm.As a result, the feature vectors of each speaker were classifted into k clusters, and the vectors in each cluster had the most similarity in both distance and variation trend, but separated those in other clusters.Moreover, a new speaker verification system was established by using GMM model for analyzing and matching the joint probability in the new feature classification space.The experiments with the standard TIMIT databases and NIST databases were implemented, and the results showed the proposed algorithm achieved good classification performance.