面向情感语音识别的情感维度PAD预测

针对现有的情感特征仅从信号的角度对情感进行分析，不能直观反映情感状态的问题，提出将连续情感维度PAD引入情感识别. 实验样本选用TYUT2.0数据库和柏林语音库中的3种情感（悲伤、愤怒和高兴），提取情感特征（韵律特征、共振峰、MFCC和非线性特征）. 为了获取客观、精确的PAD维度，利用灰色关联分析（GRA）选取影响P、A、D的主要特征，通过主成分分析（PCA）提取主要特征的主成分，将主成分作为最小二乘支持向量机（LSSVM）的输入预测P、A、D. 分别对情感特征、PAD维度及它们的融合，采用支持向量机进行情感识别. 实验结果表明，该预测方法在一定程度上提高了对P、A、D的预测精度，预测值可以有效识别情感，对情感特征在情感识别方面有一定的补充作用.

Prediction of emotional dimensions PAD for emotional speech recognition

The continuous emotional dimension PAD (pleasure, arousal, dominance) was proposed to introduce into emotion recognition in view of the fact that the existing emotional characteristics only analyze emotion from the point of view of signal, and can not directly reflect the emotional state. The experimental samples were based on three emotions (sadness, anger and happiness) from the TYUT2.0 database and the Berlin voice library, and the emotional features (prosodic feature, formant, MFCC and nonlinear feature) were extracted. Grey relational analysis (GRA) was used to select the main features that affect P, A and D in order to obtain the objective and accurate PAD dimension values. Then principal component analysis (PCA) was used to extract the principal components of the main features, and was made as the input of least squares support vector machine (LSSVM) to predict the P, A and D. The emotional features, PAD dimensions and their fusion were used separately for emotion recognition by using support vector machine. The experimental results show that the prediction method improves the prediction accuracy of the P, A and D to a certain extent. The predictive values can effectively identify the emotion, which has a certain complement to emotional characteristics in emotion recognition.