摘要
为解决网络训练复杂度高的问题并改进语音情感特征提取,提出了基于双嵌套残差网络(DNResNet11)与通道注意残差网络(CRNet)的双支路特征提取模型。首先,设计了低复杂度的DN-ResNet11以高效提取语谱图的融合情感特征,提升情感识别率;然后,结合多尺度引导滤波和局部二值模式(local binary pattern,LBP)算法对语谱图进行细节增强;最后,融合两组特征进行情感分类,形成双支路加权融合模型(weighted fusion model based on dual nested residual and channel residual network,WFDN_CRNet),进一步提升情感表征能力。在CASIA、EMO-DB、IEMOCAP等语音情感数据集上情感识别率分别达到94.58%、85.59%、65.72%,所提方法在情感识别率优于ResNet18等基准方法的同时,显著降低了计算成本,验证了模型的有效性。
To address the high complexity of network training and improve speech emotion feature extraction,a dualbranch feature extraction model based on DN-ResNet11 and a channel attention residual network(CRNet)was proposed.Firstly,the low-complexity DN-ResNet11 was designed to efficiently extract fused emotional features from spectrograms,enhancing emotion recognition accuracy.Secondly,multiscale guided filtering and the local binary pattern(LBP)algorithm were incorporated to enhance spectrogram details.Finally,the two sets of features were fused for emotion classification,forming a dual-branch weighted fusion model(weighted fusion model based on dual nested residual and channel residual network,WFDN_CRNet),further enhancing emotional representation ability.Experiments on the CASIA,EMO-DB,and IEMOCAP speech emotion datasets show emotion recognition rates of 94.58%,85.59%,and 65.72%,respectively.The proposed method not only achieves superior emotion recognition rates compared to baseline models such as ResNet18,but also significantly reduces computational cost,demonstrating the model’s effectiveness.
作者
应娜
邹雨鉴
杨雪滢
孙文胜
叶学义
蒋银河
YING Na;ZOU Yujian;YANG Xueying;SUN Wensheng;YE Xueyi;JIANG Yinhe(School of Communication Engineering,Hangzhou Dianzi University,Hangzhou 310018,China)
出处
《电信科学》
北大核心
2025年第6期139-153,共15页
Telecommunications Science
基金
浙江省科技计划项目(No.LGF21F010003)
浙江省“尖兵”“领雁”项目(No.2022C03065)。
关键词
情感识别
双嵌套残差网络
细节增强
加权融合
emotion recognition
dual nested residual network
detail enhancement
weighted fusion
