基于GMM托肯配比相似度校正得分的说话人识别

Score regulation based on GMM token ratio similarity for speaker recognition

该文提出一种基于Gauss混合模型(GMM)托肯配比相似度校正得分(GMM token ratio similarity based score regulation,GTRSR)的说话人识别方法。基于GMM-UBM(通用背景模型)识别框架,在自适应训练和测试阶段计算并保存自适应训练语句和测试语句在UBM上使特征帧得分最高的Gauss分量编号(GMM token)出现的比例(配比),然后在测试阶段计算测试语句和自适应训练语句的GMM托肯分布的配比的相似度GTRS,当GTRS小于某阈值时对测试得分乘以一个惩罚因子,将结果作为测试语句的最终得分。在MASC数据库上进行的实验表明,该方法能够使系统识别性能有一定的提升。

A GMM token ratio similarity based score regulation approach for speaker recognition is presented in this paper to judge the reliability of a test score based on the GMM token ratio similarity. In the GMM-UBM （universal background model） method, the GMM token which is the index of the UBM component giving the highest score is saved for each frame to form a vector called the GMM token ratio （GTR） of an utterance during the training and testing phases. In the test phase, the test utterance GTR is compared to the training utterance GTR to compute the similarity for a target speaker. When the similarity is less than a threshold, the original likelihood score is regulated by multiplying by a penalty factor as the final score of this test utterance. Tests on MASC show that this method improves the speaker recognition performance.