基于改进注意力机制的多路卷积课堂语音情感识别模型  

Multi-channel Convolution Classroom Speech Emotion Recognition Model Based on Improved Attention Mechanism

作　　者：梁科晋 张海军[1] LIANG Kejin;ZHANG Haijun(College of Computer Science and Technology,Xinjiang Normal University,Urumqi 830054)

机构地区：[1]新疆师范大学计算机科学技术学院,乌鲁木齐830054

出　　处：《计算机与数字工程》2024年第9期2645-2650,共6页Computer & Digital Engineering

基　　金：新疆维吾尔自治区创新环境建设专项(人才专项计划天山雪松计划)“面向高校课堂的多模态数据情感倾向性分析的关键技术研究”(编号:2019XS08);国家自然科学基金—新疆联合基金重点项目(编号:U1703261)资助。

摘　　要：针对语音情感识别研究中增加网络的深度和宽度对识别准确率提高不明显的情况,改进了注意力机制,通过将通道注意力机制和空间注意力机制相结合,并将空间注意力机制的卷积部分改进为两层的空洞卷积,以便提取更多有价值的上下文语义信息;针对单一的情感特征无法有效表征语音情感,将多个单一情感特征进行融合,增加特征的情感表征能力。该模型在中科院自动化所汉语情感数据库(CASIA)下得到了85.24%的识别准确率,在Emo-DB数据集上得到86.58%的识别准确率,证明了模型的有效性。针对真实的课堂语音数据,该模型在实验中召回率、F1值和准确率分别达到77.77%、80.76%、79.24%,体现了较好的实用性。Aiming at the situation that increasing the depth and width of the network does not significantly improve the recognition accuracy in speech emotion recognition research,the dual-channel attention mechanism is improved.By combining the channel attention mechanism and the spatial attention mechanism,the spatial attention mechanism is combined.The convolution part is improved to a two-layer convolution in order to extract more valuable contextual semantic information.For a single emotional feature cannot effectively represent speech emotion,multiple single emotional features are fused to increase the emotional representation ability of the feature.The model obtained a recognition accuracy of 85.24% under the Chinese Affective Database of the Institute of Automation,Chinese Academy of Sciences(CASIA),and a recognition accuracy of 86.58% on the Emo-DB dataset,proving the effectiveness of the model.For the real classroom speech data,recall rate,F1 value and accuracy rate of the model in the experiment reached 77.77%,80.76%,and 79.24%,respectively,reflecting good practicability.

关 键 词：情感识别 深度学习 语音情感识别 神经网络 不均衡数据集 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术]