基于MFCC特征提取和改进SVM的语音情感数据挖掘分类识别方法研究  

SPEECH EMOTION DATA MINING CLASSIFICATION AND RECOGNITION METHOD BASED ON MFCC FEATURE EXTRACTION AND IMPROVED SVM

作　　者：张钰莎 蒋盛益[2] Zhang Yusha;Jiang Shengyi(School of Electronic Information,Hunan Institute of Information Technology,Changsha 410151,Hunan,China;School of Information Science and Technology,Guangdong University of Foreign Studies,Guangzhou 510006,Guangdong,China)

机构地区：[1]湖南信息学院电子信息学院,湖南长沙410151 [2]广东外语外贸大学信息学院,广东广州510006

出　　处：《计算机应用与软件》2020年第8期160-165,212,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61572145);湖南省自然科学基金项目(2020JJ5397)。

摘　　要：设计一种语音情感数据挖掘分类识别方法。对语音情感信号进行预处理,进一步从语音话语中提取Mel频率倒谱系数(MFCC)和Mel能谱动态系数(MEDC);使用支持向量机(SVM)来分类不同的情绪状态,如愤怒、快乐、悲伤、中立、恐惧等,并基于径向基函数(RBF)内核进行训练阶段;应用柏林情感数据库和CASIA汉语情感语料库从情绪语音文件中提取特征。实验结果表明,柏林数据库和CASIA汉语情感语料库的正确识别率分别为82%和90.39%。与几种先进的对比方法进行比较,该方法在不同降维、不同信噪比下均取得了最优的识别精度。We design a speech emotion data mining classification and recognition method.The speech emotion signal was preprocessed,and the Mel frequency cepstral coefficient(MFCC)and the Mel energy spectrum dynamic coefficient(MEDC)were extracted from the speech discourse;SVM was used to classify different emotional states,such as anger,happiness,sadness,neutrality,fear,etc.,and the training phase is based on a radial basis function(RBF)kernel;the Berlin Emotion Database and the CASIA Chinese emotional corpus were used to extract features from emotional speech files.The experimental results show that the correct recognition rates of the Berlin database and the CASIA Chinese emotional corpus are 82% and 90.39%,respectively.Compared with several advanced comparison methods,our method achieves the best recognition accuracy under different dimensionality reduction and different SNR.

关 键 词：语音情感识别 支持向量机 数据挖掘 MEL频率倒谱系数 Mel能谱动态系数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]