Research On Chinese Tone Recognition Based On Surface Electromyographic Signals Of Human Body

While the current speech recognition of the Chinese language in low signal-to-noise environments is insufficient,Chinese voice tone recognition is of enormous importance in semantic comprehension,speech recognition,speech synthesis,and other application areas.Utilizing the surface EMG signal for voice tone detection in environments with significant background noise is an efficient technique to increase the recognition rate because it is insensitive to environmental noise.A method based on surface EMG signal for speech tone identification is created from Chinese voice tone recognition in order to test this theory.The following areas were the main research areas:(1)Based on the time-frequency characteristics of surface EMG signals,the MFCC feature parameters were improved to enhance the expression of low-frequency information of surface EMG signals in the range of 0-1000 Hz.Comparison experiments were designed and validated in three traditional machine learning classification algorithms such as SVM,KNN and RF,and two classification algorithms based on temporal features such as RNN and LSTM,respectively.(2)A feature fusion method is proposed to fuse MFCC features with short-time time-domain features for the lack of information of MFCC feature parameters in the time domain.Eleven short-time time-domain features commonly used in surface EMG signal recognition are extracted,and five feature selection algorithms are used to filter and combine these 11 time-domain features,and 26 different feature fusion schemes are proposed,from which five fused features with higher accuracy are selected by LSTM classification algorithm.(3)LSTM,GRU and BiLSTM network models are improved by introducing one-dimensional convolutional and pooling layers,and different combinations of features are identified for classification.By comparing the recognition results of various models on various features,it is found that the optimal performance is achieved by using STD,Kurtosis,ShapeFactor,ImpulseFactor + MFCC,a fusion feature as input,trained on the 1DCNN-BiLSTM network model.The experiments show that the improved MFCC features + STD,Kurtosis,ShapeFactor,ImpulseFactor are used to form fusion features on the 1DCNN-BiLSTM network model for the Chinese vowels "a","o","e" on the 1DCNN-BiLSTM network model,with the accuracy of 94.6%,95.1%,and 92.3%,respectively,which is more than 20% higher than the traditional MFCC+LSTM method,and can effectively perform the task of Chinese tone recognition.