Speech Fatigue Detection Based On Deep Learning

Station attendants are responsible for coordinating all aspects of railway transportation,such as timely response to equipment failure,provisional speed limit,plan adjustment and other emergencies.High-intensity work for a long time would bring physical and psychological fatigue to station attendants,and may cause immeasurable harm to railway safety.Currently,only a measure of two-hour rest in their position has been taken to ensure the staffs’ ability to work,which is not enough.To cope with this risk,it is of great research significance to track,detect and evaluate the work fatigue status of the station duty staff in real time.Voice carries a large amount of information on physical and mental fatigue.A lot of voice call responses are used in the work of station attendants.In order to know the fatigue information of the station attendants and to reduce the potential safety hazards caused by it.Combining the currently popular learning method--deep learning,this thesis judges the fatigue information of the station attendants through the speech recognition by detecting the voice of the station attendants.The main work of this thesis is as follows:(1)Collection of voice data samples to build a fatigue voice database.This thesis defines three states of human fatigue: normal state,fatigue state,and very fatigue state,and designs 300 sentences on duty for daily use.15 volunteers with accents in different regions of China recorded the sample,and their ages ranged from 20 to 45,including 2 femal and 13 male.Manually recorded and marked 5,033 voice samples to build a train attendant fatigue detection voice database(TA-FD).(2)Establish an Attention-based Convolutional Recurrent Neural Network(ACRNN)model and conduct a experiment to recognize fatigue speech.The network model is composed of Convolutional Neural Network(CNN),Long Short Term Memory Network(LSTM),and Attention Mechanism.The experimental results show that the accuracy of the ACRNN model on the TA-FD dataset is 89.67%.Compared with the CNN and Convolutional Recurrent Neural Network(CRNN)models,the recognition accuracy of the ACRNN model is improved by 5.49% and 3.74%,respectively.(3)To further improve the accuracy of fatigue speech detection.Convolutional Recurrent Neural Network augmented with Capsule layer(CRNN-Cap)is proposed,which is suitable for fatigue speech classification.The network architecture consists of CNN,LSTM and capsule network.This completes the feature extraction of the input information.The final results are divided into three fatigue states.Experimental simulation results show that the accuracy rate of the CRNN-Cap model on the TA-FD dataset is 93.44%,which is 3.77% higher than the classification accuracy of the ACRNN model.