基于3D倒谱特征和注意力机制的CNN-BiLSTM语音情感识别  

CNN-BiLSTM with 3D Cepstrum Feature and Attention Model for Speech Emotion Recognition

作　　者：夏玉果[1] 杜静 赵力[2] XIA Yuguo;DU Jing;ZHAO Li(School of Microelectronics,Jiangsu Vocational College of Information Technology,Wuxi Jiangsu 214153,China;School of Information Science and Engineering,Southeast University,Nanjing Jiangsu 210096,China)

机构地区：[1]江苏信息职业技术学院微电子学院,江苏无锡214153 [2]东南大学信息科学与工程学院,江苏南京210096

出　　处：《电子器件》2022年第6期1434-1440,共7页Chinese Journal of Electron Devices

基　　金：国家自然基金重点项目(U2003207);江苏省高等学校自然科学研究项目(19KJB510027);2021年江苏省高校优秀科技创新团队项目;2021年江苏省高职院校教师专业带头人高端研修项目;2021年中国高校产学研创新基金项目(2021ITA05016)。

摘　　要：为了进一步提升语音情感特征的区别性,提出一种基于3D倒谱特征和注意力机制的语音情感识别方法。在提取典型特征MFCC的基础上,融合其一阶和二阶差分形成动态的3D倒谱特征矢量,然后利用卷积神经网络和双向长短期记忆网络搭建双向循环卷积神经网络(CNN-BiLSTM)进行长时建模,并利用注意力机制,进一步增强语音关键性情感特征的权重。实验结果表明,该方法能聚焦特征的有用信息,从而有效提高语音情感识别的准确率,在汉语情感语料库CASIA上的的准确率达到90.48%。In order to further enhance the discrimination of speech emotion feature,an approach for speech emotion recognition using convolutional neural network-bi-directional long short-term memory(CNN-BiLSTM)based on 3D cepstrum feature and Attention Model is proposed.On the basis of extracting typical feature MFCC,the 3D cepstrum feature is constructed by fusing its first-order and second-order difference to form dynamic input feature vector.Then CNN-BiLSTM is constructed using convolutional neural network and bidirectional long short-term memory network for long-term modeling,and attention mechanism is utilized to further enhance the weight of key emotional features of speech.The experimental results show the proposed method can focus on the useful information of the feature and effectively improving the accuracy of speech emotion recognition.The accuracy of the proposed method on the data set of Chinese emotion corpus CASIA reaches 90.48%.

关 键 词：语音情感识别 3D谱特征 注意力机制 双向长短期记忆网络 

分 类 号：TN912.3[电子电信—通信与信息系统]