基于SVM-HMM混合模型的说话人确认

Speaker Verification With the Hybrid Use of Support Vector Machine and Hidden Markov Model

提出一个文本无关的说话人确认的算法 .该算法将支持向量机 (SVM)的输出通过 Sigm oid函数和高斯模型转化为概率 ,并作为隐式马尔可夫模型 (HMM)中各个隐状态的输出概率 .由于 HMM适于处理连续信号 ,SVM适于处理分类问题 ;同时 ,HMM更多地表达了类别内部的相似性 ,而 SVM则很大程度上反映了类别间的差异 ,因而根据两者不同的侧重点 ,使其组合获得了很好的效果 .

HMM is good at dealing with sequential inputs, while SVM shows superior performance in classification. Furthermore, the former approach usually provides an intra-class measure while the latter proposes inter-class difference. Since these two classifiers use different criteria, they can be combined to yield an ideal one. The output of support vector machines is converted into the form of posterior probability which is computed by the combined use of sigmoid function and Gaussian model, it acts as a probability evaluator in the hidden states of hidden Markov models. Experiments on speaker verification show that this hybrid model achieves better performance than continuous density hidden Markov models.