Research On Speech Recognition Based On The Degree Of Fatigue

With the rapid development of computer technology, Information technology based on computer technology will have a more profound impact on human society. Early on, the information processed by computer is usually number, text, sound, video. Although there is no change in the methods of human information expression, but the depth of information processing of the computer has far exceeded.The early speech signal processing mainly contains the recognition of semanteme and speech synthesis, which has been the primary research direction.With the further study of the speech recognition, the speech signal processing has been refined.This thesis focuses on the research of feature extraction of fatigue characteristic parameters contained in the speech and the classification of the voice fatigue degree using the BP neural network and support vector machine. The main research work is shown as follows:(1) Pre-processing of speech signal. In this part, pre-processing has been carried out on the speech signal. Firstly, the analog speech signal has been digitized through the voice signal pick-up device. Then, pre-emphasis and framing and endpoint detection are carried out in succession to acquire the subsequent processing data. Besides, the thesis has compared the windowing effect of Rectangular window and Hann Window and Hamming window.(2) The classification of speech fatigue level. There is still no unified classification criteria due to the immature research on the speech fatigue degree, so the thesis defines four categories of speech fatigue degree:non fatigue, mild fatigue, severe fatigue and extreme fatigue degree. The related parameters of speech data in different fatigue degrees have been analyzed and compared in MATLAB. Sound vibration, pitch frequency and resonance peak of speech signal in high fatigue degree have changed obviously. So 11 characteristic parameters including the above three objects have been extracted to provide the necessary support for the recognition of fatigue degree.(3) Classification of fatigue characteristic parameters based on BP (Back Propagation) neural network and SVM (Support Vector Machine).Firstly, the theory of BP (Back Propagation) neural network has been described. Then the weight of BP network has been calculated using the gradient algorithm. Besides, the SVM (Support Vector Machine) classifier has been introduced and the optimal hyperplane has been established. Finally, a series of simulation experiments have been carried out in MATLAB.(4) Realization of speech fatigue recognition software. A number of speech signal in different fatigue degrees have been used to train based on BP and SVM. The above experiments have been realized in the environment of VC++. Besides, the system interface of the software has been designed, which add the function of displaying the speech energy to observe the fatigue state of speech signal easily. Qt can be used to program if the software of speech fatigue recognition is needed in the cases of crossing platform. Finally, the test experiments on the speech fatigue recognition software are carried out. The experimental results reveal that the speech fatigue recognition system has good robustness in different fatigue degrees.