基于语谱图和卷积神经网络的语音情感识别

Speech emotion recognition based on spectrogram and CNNs

针对语音情感识别的特征提取和分类模型构建问题,首先提出了一种基于语谱图的特征提取方法,将语谱图进行归一灰度化后,利用Gabor滤波器进行纹理特征提取,并采用主成分分析(principal component analysis,PCA)对特征矩阵进行降维;然后分析了卷积神经网络(convolutional neural networks,CNNs)并把其作为情感识别分类器;最后在Emo DB和CASIA库进行了不同的比对实验.实验结果取得了较高情感识别率,表明了所提特征提取方法的有效性以及CNNs用作情感分类的可行性.

To solve the problem of feature extraction and classification in speech emotion recognition,first a feature extraction method based on spectrogram was proposed,the method uses Gabor filter to extract the texture feature from the normalized spectrum gray image,and reduce these feature matrix dimension using the PCA.Then the convolutional neural networks was used as an emotion recognition classifier.Finally the performance of this system was assessed by computer simulations and a higher recognition rates were achieved respectively on the Emo DB and CASIA database through comparative experiment in different conditions,the results showed that the method proposed in this paper is effective and the CNNs can be used successfully for emotion recognition as a classifier.