针对训练样本与测试样本来自不同语音情感数据库造成特征向量空间分布不匹配的问题,采用半监督判别分析减小二者的差异。首先寻找有标签的训练样本和来自另一个库的部分无标签训练样本之间的最优投影方向。基于一致性假设即相近的点更有可能具有相同的类别,利用p近邻图对无标签训练样本相近点之间的关系进行建模,从而获得无标签样本的分布信息。在保证无标签样本间流形结构的同时,使所有训练样本类间散度和类内散度的比值达到最大,从而得到最优的投影方向。采用两组实验进行验证,第1组用eNTERFACE库训练去测试Berlin库,识别率为51.41%,第2组用Berlin库训练测试eNTERFACE库,识别率为45.76%,相比未采用半监督判别分析的识别结果分别有了13.72%和22.81%的提高,说明该算法的有效性。通过实验前后数据的可视化分析,说明利用半监督判别分析确实减小了不同库之间特征向量空间分布的不匹配问题,从而提高跨库语音情感识别率。 Aiming at the mismatch of spatial distribution of feature vectors caused by different speech emotion databases from training samples and test samples, semi-supervised discriminant analysis is adopted to reduce the difference between the two. First, look for the optimal projection direction between a labeled training sample and a partially unlabeled training sample from another library. Based on the consistency hypothesis that similar points are more likely to have the same category, the p-nearest neighbor graph is used to model the relationship between similar points in unlabeled training samples to obtain unlabeled sample distribution information. While ensuring the manifold structure between the unlabeled samples, the ratio of the divergences and the divergences of all kinds of training samples is maximized so as to obtain the optimal projection direction. Two groups of experiments were used to verify that in the first group, Berlin library was tested by eNTERFACE library, the recognition rate was 51.41%. In the second group, the eNTERFACE library was trained and trained with Berlin library, the recognition rate was 45.76%. Compared with the semi-supervised discriminant analysis The recognition results have been improved by 13.72% and 22.81% respectively, which shows the effectiveness of the algorithm. Through the visual analysis of the data before and after the experiment, it shows that the semi-supervised discriminant analysis does reduce the problem of mismatch of the spatial distribution of eigenvectors among different databases, so as to improve the emotion recognition rate of cross-bank speech.