基于两种GMM-UBM多维概率输出的SVM语音情感识别

摘要

针对GMM应用于情感识别时区分能力较弱的缺点,提出了一种将GMM与SVM有效结合的算法:基于GMM-UBM多维概率输出的SVM语音情感识别方法.该方法将GMM-UBM模型对一条语音的情感特征参数的两种多维概率输出(与特征向量同维、与GMM阶数同维)作为SVM分类器的特征参数,既利用了GMM表征数据本身统计特性的能力,又保留了SVM判决能力强的特点.在柏林情感语音库与汉语情感语料库上进行的实验结果表明,该方法在语音情感识别上的平均识别率较标准GMM方法提高1.7%～3.7%.%Aiming at the poor ability of discrimination in the case of speech emotion recognition by using GMM model,this paper proposed an algorithm based on GMM-UBM multidimensional likelihoods and SVM, which combined the advantages of both GMM and SVM. It regarded the two kind of GMM-UBM multidimensional likelihoods( the same dimension with eigenvector and mixtures of GMM ) for one test speech as feature of emotion for SVM. This method took advantage of the statistical properties of characterization of GMM and the strong discrimination ability of SVM. Experimental results on Berlin emotional speech databases and emotional speech databases Mandarin demonstrate that the proposed method achieves significant improvements about 1.7％ to 3.7％ than standard GMM on speech emotion recognition.