基于分类特征映射的SVM话者确认  

Speaker Verification Based on Pitch Classified Feature Mapping

作　　者：贺庆玮[1] 李辉[1] 许敏强[1] 

机构地区：[1]中国科学技术大学电子科学与技术系,合肥230027

出　　处：《通信技术》2010年第3期147-149,共3页Communications Technology

摘　　要：为了解决与文中无关的话者确认,大量训练样本数据所导致的建立支持向量机SVM(SupportVectorMachine)话者模型困难,文中提出了一种基于基音分类特征映射和支持向量机的话者确认系统,首先根据基音周期将语音倒谱参数在特征空间上分类,再利用GMM-UBM结构进行特征映射,获得每个特征子空间中的话者特征参数并建立SVM话者模型。基音分类特征映射不仅使得样本数据极大地压缩,而且让子空间中SVM分类界面具有更好的区分性,因此,对各分类子系统评分融合之后的总系统具有更好话者确认性能。在NIST’06数据库上的实验证明了该方法的有效性。In order to solve the problem of how to train the SVM speaker model properly while giving large quantities of speech features in text-independent speaker verification,an approach based on pitch classified GMM-UBM feature mapping and SVM is proposed,which classifies speech in the feature space by the pitch parameter and then carries out feature mapping by using GMM-UBM structure,thus to get the speaker feature and build SVM speaker models in each feature sub-space.The pitch classified feature mapping not only condenses the quantity of speech cepstral feature,but also optimizes the discriminability of SVM.Therefore,after linearly fusing the scores of sub-systems,a better speaker verification performance is obtained.The experiments in the database of NIST＇06 indicate that this approach is fairly efficient.

关 键 词：分类 基音周期 高斯混合模型 特征映射 支持向量机 话者确认 

分 类 号：TN912.3[电子电信—通信与信息系统]