基于CGRU模型的语音情感识别研究与实现  

Research and Implementation of Speech Emotion Recognition Based on CGRU Model

作　　者：郑艳[1] 陈家楠 吴凡 付彬 ZHENG Yan;CHEN Jia-nan;WU Fan;FU Bin(School of Information Science&Engineering,Northeastern University,Shenyang 110819,China)

机构地区：[1]东北大学信息科学与工程学院,辽宁沈阳110819

出　　处：《东北大学学报（自然科学版）》2020年第12期1680-1685,共6页Journal of Northeastern University(Natural Science)

基　　金：国家自然科学基金资助项目(61773108).

摘　　要：目前普遍使用深度神经网络用于语音情感特征的提取,但使用哪种神经网络模型、如何缓解模型过拟合问题还需进一步研究.针对这些问题,提出了一种结合一维卷积(CNN)以及门控循环单元(GRU)的CGRU模型,从原始语音信号的MFCC特征中提取语音的低阶以及高阶情感特征,并通过随机森林对其进行特征选择,在三种公用的情感语料库EMODB,SAVEE,RAVDESS上分别取得了79%,69%以及75%的识别精度.通过添加高斯噪声及改变速度等方法来增加样本量实现数据扩充,进一步提高了识别精度.通过在线识别系统验证了模型在实际环境中的可用性.Speech emotion recognition is a very important research direction in emotion computing and human-computer interaction.At present,deep neural network is widely used to extract emotional features of speech,but further research is needed on which neural network model to use and how to alleviate the problem of model overfitting.To solve these problems,a CGRU model was proposed,which combined one dimensional convolutional neural networks(CNN)and gated circulation unit(GRU).The low-order and high-order emotional features of speech were extracted from the MFCC features of the original speech signal,and the features were selected through random forest,which achieved 79%,69%and 75%recognition accuracy respectively on three common emotional corpus:EMODB,SAVEE,RAVDESS.By using the data augmentation technique,the sample size was increased by adding gaussian noise and changing the speed,which further improved the identification accuracy.The availability of the model in the real world was verified through the online identification system.

关 键 词：语音情感识别 梅尔频率倒谱系数 CGRU模型 随机森林 数据扩充 

分 类 号：TN912.3[电子电信—通信与信息系统]