仿选择性注意机制的语音情感识别算法  

Speech emotion recognition algorithm based on the selective attention mechanism

作　　者：梁瑞宇[1,2] 赵力[2] 陶华伟[2] 王青云[1,2] 邹采荣[2] 

机构地区：[1]南京工程学院通信工程学院,南京211167 [2]东南大学信息科学工程学院,南京210096

出　　处：《声学学报》2016年第4期537-544,共8页Acta Acustica

基　　金：国家自然科学基金(61301219;61375028);江苏省自然科学基金(BK20130241);江苏省"青蓝工程"和南京工程学院高层次引进人才科研启动基金(YKJ201526)资助

摘　　要：有效特征的选取一直都是语音情感识别算法的关键。为此,针对语音情感特征选择与构建的问题,一种仿选择性注意机制的语音情感识别算法被提出。考虑到语音信号的时频特性,算法首先计算语音信号的语谱图;其次,模仿选择性注意机制,计算语谱图的颜色、方向和亮度特征图,归一化后形成特征矩阵;然后,将特征矩阵重排列并进行PCA降维,形成情感识别特征向量;最后,利用改进的支持向量机分类方法进行语音情感识别。对愤怒、恐惧、高兴、悲伤和惊奇5种情感的识别实验显示,基于选择性注意的方法能够获得较好的识别效果,平均识别率为85.44%。相比于韵律特征和音质特征,语音情感识别率至少提高10%;相比于其它语谱特征,识别率提高7%左右。The selection of effective features is always the key issue of speech emotion recognition algorithm. For the problem of the speech emotion features selection and construction, one speech emotion recognition algorithm based on the selective attention mechanism is proposed. Firstly, considering the time-frequency of speech, the spectrogram is calculated. Secondly, imitating the selective attention mechanism, the color, direction and brightness map of the spec- trogram is computed. Based on these maps, the feature matrix is formed after normalization. Then, the feature matrix is rearranged and reduced the dimensionality by the principal component analysis （PCA） to form the emotion recognition feature vector. Finally, the speech emotion is classified with the improved support vector machine （SVM）. Experiments on emotion （anger, fear, joy, sadness and surprise） recognition show the proposed method has an improved recognition rate 85.44%. Compared with the prosody and voice quality features, the speech emotion recognition rate has been improved at least by 10%. And compared with others spectrogram features, the improved recognition rate reaches 7%.

关 键 词：语音情感识别 选择性注意 语音信号 特征矩阵 识别实验 特征图 算法类 颜色信息 特征提取 情感特征 

分 类 号：TN912.34[电子电信—通信与信息系统]