摘要
在通过陆空通话语音判别管制员疲劳的任务中,根据疲劳语音特点针对性构建特征,可以有效地提升判断准确性。选择语音情感识别任务中传统的韵律特征为基础,结合对疲劳语音的分析建模,并与传统特征一起重新建立一套音节级的特征组合方案。在适用于kNN、SVM等传统分类的同时,在结构上还适应了时间序列性的学习方法。实验结果表明,特征方案在使用使用常见的分类算法时,比传统特征可有20%~55%的性能提升,证明了新引入的特征与语音疲劳任务的相关性,并可有效提升疲劳判别准确性。
This paper presents a feature extraction method used for speech fatigue recognition of ATC from radiotelephony datas.Our features are modeled referring the descriptive characteristics of radio telephony datas that considered to express fatigue,and the features are founded on tradional prosody,articulation and speech quality features with a syllable-level structure,which adapts to not only the traditional classifiers as kNN,SVM,but also the method for time series as HMM,RNN.Experimental results show that the recognition rate improved up to 55%compared to traditional features,which prove the validity of our feature in task of speech fatigue recognition of ATC.
作者
李兆悦
LI Zhao-yue(Tianjin Air Traffic Management Operational Planning and Safety Technology Laboratory,Air Traffic Management Research Base,Civil Aviation University of China,Tianjin 300300,China)
出处
《航空计算技术》
2020年第5期56-60,共5页
Aeronautical Computing Technique
基金
国家重点研发计划项目资助(2016YFB0502401)
民航华东空管局科技项目资助(KJ1804)。
关键词
空中交通管制
陆空通话
精神疲劳
模式识别
声学特征
air traffic control
radiotelephony communication
mental fatigue
pattern recognition
acoustic feature
作者简介
李兆悦(1994-),男,北京市人,硕士研究生,主要研究方向为管制员疲劳、语音疲劳判别。
