摘要
传统语音情感识别主要基于单一情感数据库进行训练与测试。而实际情况中,训练语句和测试语句往往来源于不同的数据库,识别率较低。为此,该文提出一种基于子空间学习和特征选择融合的语音情感识别方法。通过采用回归方法来学习特征的子空间表示;同时,引入l2,1-范数用于特征的选择和最大均值差异（maximum mean discrepancy,MMD）来减少不同情感数据库间的特征差异,进行联合优化求解从而提取较为鲁棒的情感特征表示。在EMO-DB和eNTERFACE这2个公开情感数据库上进行实验评价,结果表明：该方法在跨库条件下具有较好的性能,比其他经典的迁移学习方法更加鲁棒高效。
Traditional speech emotion recognition methods are trained and evaluated on a single corpus.However,when the training and testing use different corpora,the recognition performance drops drastically.A joint subspace learning and feature selection method is presented here to imprive recognition.In this method,the feature subspace is learned via a regression algorithm with the l2,1-norm used for feature selection.The maximum mean discrepancy（MMD）is then used to measure the feature divergence between different corpora.Tests show this algorithm gives satisfactory results for cross-corpus speech emotion recognition and is more robust and efficient than state-of-the-art transfer learning methods.
作者
宋鹏
郑文明
赵力
SONG Peng;ZHENG Wenming;ZHAO Li(School of Computer and Control Engineering Yantai University, Yantai 264005, China;Key Laboratory of Child Development and Learning Science of Ministry of Education, Southeast University, Nanjing 210096, China)
出处
《清华大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2018年第4期347-351,共5页
Journal of Tsinghua University(Science and Technology)
基金
国家自然科学基金资助项目(61703360,61602399)
山东省自然科学基金资助项目(ZR2014FQ016,ZR2016FB22,ZR2017QF006)
东南大学基本科研业务费资助项目(CDLS-2017-02)
关键词
特征选择
子空间学习
情感识别
feature selection
subspace learning
emotion recognition
