论文部分内容阅读
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 speaker identification with SVM (Support Vector Machine Machine) caused by extensive training of sample data, this paper proposes a speaker recognition system based on pitch classification feature mapping and SVM. Firstly, The pitch period classifies the speech cepstrum parameters in the feature space, and then uses the GMM-UBM structure to perform feature mapping to obtain the speaker feature parameters in each feature subspace and establish the SVM speaker model. Pitch classification feature mapping not only makes the sample data greatly compressed, but also makes the SVM classification interface in sub-space have better distinguishability. Therefore, the overall system after the fusion of the classification subsystems has better speaker verification performance. Experiments on the NIST’06 database demonstrate the effectiveness of this method.