Speech Fatigue Detection Based On Neural Network

Safety is vital in railway transportation.Station attendant is an important post closely related to transportation safety.The fatigue of station attendant will affect work efficiency and even transportation safety.Considering that station attendant needs frequent verbal instructions in their work,the fatigue state of station attendant can be monitored in real time through voice,and the hidden dangers to the safety of railway transportation brought by the fatigue of station attendant can be reduced effectively.In this study,speech samples with different fatigue states are collected,fatigue features of speech signal are extracted,the relationship between different speech features and fatigue states is explored,and the detection of speech fatigue based on selected speech features is realized through neural network.The main research work in this study is shown as follows:First,collection of speech data.Currently,there is a lack of public fatigue-related speech data sets.In this study,self-collected speech data from male and famale is used,two kinds of speech contents are included,and the speech data is labeled as four fatigue states:very spiritual,normal,relative fatigue and very tired.Second,fatigue features extraction of speech signal.In this study,short-term energy,short-term average zero-crossing rate,Mel frequency cepstrum coefficients,fundamental frequency and first formant are selected to analyze the speech fatigue states.Firstly,speech signal is enframed and denoised,then the endpoint is detected according to the energy entropy ratio,and finally corresponding algorithm is used to extract the features.Third,analysis and process of feature data.First,outliers in the feature data are eliminated to improve data quality,then fatigue relevance analysis of speech features for single-content speech,different-content speech and different-content speech of different speakers is carried out respectively,finally,the feature parameter group used in fatigue classification experiment is determined.Fourth,the experiment of speech fatigue classification based on neural network.A neural network is built with Tensorflow,and it is trained with Adam algorithm based on processed speech feature data,then the accuracy of fatigue discrimination with speech features is tested.Concluding the results of several experiments,the discrimination of speech fatigue on the data set used in this study have made good results.