改进SVM结合决策树的情感语音识别

Speech emotion recognition using decision tree and improved SVM

针对传统情感语音识别方法整体分类精度不高的问题,提出改进SVM结合决策树的情感语音识别方法。首先,提取能量、过零率、幅度以及线性预测系数共四个时域统计特征,基于互相关技术将中性情感语音作为参考,与其余的情感语音相关联。其次,从每个得到的互相关序列中提取出质心等五个特征。最后,利用提出的改进SVM与决策树的混合模型完成分类识别。在柏林情感语音数据库上的结果表明,提出的方法能有效完成情感语音信号识别。

Aiming at the problem that the overall classification accuracy of traditional emotional speech recognition methods is not high,an emotional speech recognition method is proposed based on improved SVM classifier and decision tree.Firstly,four time-domain statistical features including energy,zero-crossing rate,amplitude and linear prediction coefficients are extracted,and the neutral emotion speech are further used as a reference based on the cross-correlation technique,and is associated with the rest of the emotional speech.Then,five features such as the centroid are extracted from each of the obtained cross-correlation sequences.There is a contradiction in logic when this paragraph is reorganized.Finally,classification recognition is performed based on the proposed hybrid model of SVM classifier and decision tree.The classification accuracy of the classifier is tested by the Berlin emotional speech database.The results show that the proposed method can effectively complete the emotional speech signal recognition.