论文部分内容阅读
针对信道失配和统计模型区分性不足而导致话者确认性能下降问题,文中提出一种将因子分析信道失配补偿与支持向量机模型相结合的文本无关话者确认方法.在SVM话者模型前端采用高斯混合模型-背景模型(GMM-UBM)方法对语音特征参数进行聚类和升维,并利用因子分析(FA)方法,对聚类获得的超矢量进行信道补偿后作为基于SVM话者确认的输入特征,从而有效解决SVM用于文本无关话者确认的大样本、升维问题,以及信道失配对性能影响问题.在NIST06数据库上实验结果表明,文中方法比未做失配补偿的GMM-UBM系统、GMM-SVM系统在等误识率上有50%以上的改善,比做了FA失配补偿的GMM-UBM系统也有15.8%的改善.
Aiming at the problem of speaker recognition performance degradation due to channel mismatch and statistical model insufficiency, a text-independent speaker verification method combining factor analysis channel mismatch compensation with support vector machine model is proposed in this paper.In SVM speaker model At the front end, Gaussian mixture model-background model (GMM-UBM) is used to cluster and up-scale the speech feature parameters. The factor vector (FA) method is used to compensate the hyper- Confirm the input features, so as to effectively solve the problem of large sample, dimension up-scaling problem and channel mismatch impact on the performance of the SVM in the NIST06 database.Experimental results on the NIST06 database show that the proposed method is better than the GMM -UBM system and GMM-SVM system have more than 50% improvement in the misclassification rate and 15.8% improvement over the GMM-UBM system with FA mismatch compensation.