期刊论文

Research on speech emotion recognition based on convolution feature extraction and fusion

中文

电子测量技术

1002-7300

2018

41

16

138-142

10.19651/j.cnki.emt.1801649

湖北省科技馆国家科技支撑计划（2015BAK33B）项目资助；

本校为第一机构

语音情感识别中,情感特征的选取是决定最终识别结果的关键因素。传统的谱特征来源于语谱图特征的再加工,存在因分帧处理引起相邻谱特征相关性被忽略以及谱特征与目标标签不相关的问题,导致语谱图部分特征信息丢失。为此,提出卷积特征与谱特征融合的语音情感识别方法。选用合适的卷积神经网络（CNN）,从语谱图中自动提取隐含的、有效的情感特征,再融合语音情感统计学谱特征,构造多分类支持向量机（SVM）来识别不同的语音情感。在包含7种基础情感的柏林情感数据集上进行实验,获得86.4%的识别率,比传统谱特征方法高6.2%,提出的卷积特征提取与融合的方法能有效识别语音情感。

In speech emotion recognition, the selection of emotional characteristics is the key factor to determine the final identification result. Traditional spectral characteristic spectrogram is derived from the language characteristic of reprocessing, adjacent spectral characteristics caused by framing processing relativity are ignored and label not related problems, with the target spectrum characteristics lead to language spectra part feature information loss. To this end, the speech emotion recognition method of convolution characteristic and spe...