Extraction And Detection Of Fatigue Information In Speech

Fatigue state can reflect the change process of human physiological and psychological.Fatigue will lead to the decline of human body functions,induce diseases and even lead to safety accidents.When people speak,the speech can not only transmit language and text information,but also convey people's mental state information.If the driver's fatigue status can be monitored in real time by speaking,the risk of accidents can be effectively reduced.To this end,the speech database under different fatigue states are recorded in this paper to explore the relationship between the fatigue state and the speech signal,that is,extracts speech feature parameters,conducts statistical analysis of the data,and fuses each feature parameter into the Gaussian mixture model classifier to find the best detection performance.The specific work of this paper is as follows:(1)Collection of fatigue speech corpus.At present,there is no recognized standard fatigue speech corpus.the speech database with different fatigue degrees are recorded,including 15 testers.Each sample has four different fatigue states,corresponding to mental(sufficient sleep),general mental(forced without sleep for 12 hours),mild fatigue(forced without sleep for 24 hours),and severe fatigue(forced without sleep for 36 hours).(2)Extraction and analysis of speech feature parameters.Firstly,the speech data are preprocessed,including framing,windowing,endpoint detection,and five features are selected,i.e.,intensity,pitch,formant,sliding differential cepstrum feature and mel frequency cepstrum coefficients,the algorithm implementation principle of these feature parameters are introduced.Then feature parameter extraction is performed,and statistical analysis is performed from the extracted data to find the relation between fatigue state and each feature parameter.(3)Gaussian mixture model is used as a classifier to detect and classify speech fatigue.Firstly,the algorithm principle of Gaussian mixture model is introduced,including K-means(K-means)clustering and Expectation Maximization(EM)algorithm.Secondly,each feature parameter is sent to the classifier separately for fatigue detection,and the performance of each feature is investigated separately.Then the feature parameters are fused with each other,and the detection performance after the fusion of various features was further investigated.The experimental results show that for single-feature fatigue detection,the mel-frequency cepstral coefficients obtain the best performance,and for fusion feature,the detection performance is higher than that of single-feature.After all features are fused,the detection accuracy rate can reach 91%.(4)In order to facilitate the communication with medical personnel and make their extraction of feature parameters more convenient,a feature parameter extraction system is designed by using Labview software.The system is divided into two interfaces: single speech feature extraction and multiple speech feature extraction,which can extract seven speech feature parameters involved in this paper.The system can import the selected speech database path,input the parameter configuration,display the corresponding waveform of the feature parameters,and automatically save the characteristic parameter data into Excel format files.The system interface is friendly and clear,and the operation is simple and fast.