基于长短时特征融合的语音情感识别方法精选.pdfVIP

ISSN 1000-0054 ( ) 2008 48 S1 18/ 27 CN 11-2223/ N J T singhua Un iv ( Sci Tech) , 2008, V l. 48, N . S1 708-714 基于长短时特征融合的语音情感识别方法 韩文静, 　李海峰, 　韩纪庆 ( , 150001) : 情感识别领域中广泛使用的情感特征有基于语句的全局统计特征和基于语音帧的时序特征。针对 2 类 基于不同时长的情感特征均不能够最有效地表达情感信息的问题, 该文提出使用 “语段特征”用于识别, 并给出了 各类情感状态对应的 “最佳识别段长”。为进一步提高系统识别性能, 该文还构建了全局控制Elman 神经网络用于 将全局统计特征与基于语段的时序特征相融合。实验表明: 融合长短时特征后的平均系统识别率可达66. 0% , 与 单独使用各类特征或语段特征相比分别提高了5. 9% 和1. 7% , 同时有效降低了各情感之间的混淆度。 : 模式识别; 语音情感识别; 情感特征; Elman 神经网络 : T P 391. 4: A : 1000-0054( 2008) S1-0708-07 Speech emotion recognition with combined short and long term features HA N W enj ing, LI Haif eng, HA N Jiqing ( School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150001, China) Abstract: Utterance-based gl bal statist ics and fram e-based temp ral features have been w idely used in speech em t i n rec gniti n sy st ems, but these features can n t effectively describe all f the em ti nal inf rmati n. In this research, segment-based feat ures are ext racted and applied w ith a best segment length f r em t i n rec gniti n f r each em ti nal st ate. Further m re, a n vel neural netw rk m del named Gl bal c nt r l Elman is pr p sed t c mbine the ut terance-based feat ures and segment-based features t gether . Experiments sh w that the perf rmance f c mbined features may reach a rec gniti n rate f 66. 0% , m uch higher than bt ained by utterance-based feat ures r segment -based features. T he rec gniti n rate may be impr ved by 5. 9% and 1. 7% respectively , and the c nfusi n bet ween em ti nal state is als effectively reduced. Key words: pattern rec gniti n; speech em ti n rec gniti n; em ti n feat ure; Elman neural n