Silent Speech Recognition Method Based On High-density S Emg Information

Clear language communication is particularly important in human daily life and work,and it is the cornerstone of people's communication and interaction.However,in many cases,people inevitably have the need to receive language content indirectly.Speech-handicapped people cannot make a sound normally,and thus lose the communication bond with healthy people.For this group of people,it is urgent to rebuild new communication methods.At the same time,in scenarios such as confidential work,quiet environment,underwater work,wearing protective clothing,etc.,people have the need to exchange information without making a sound.In addition,the current popular automatic speech recognition technology has the fatal disadvantage of being susceptible to environmental noise.Combining the above three types of problems and needs,there is currently an exploration of using surface EMG signals for silent speech recognition.Surface electromyography(s EMG),as an objective physiological electrical signal detection technology,has been widely used in the field of pattern recognition,and there is no lack of exploration and research in the field of speech recognition.When performing speech recognition based on surface EMG signals,feature extraction and classifier selection are two indispensable links,which play a decisive role in the final recognition effect.Therefore,how to select and match features and classification algorithms is a great test for using EMG signals for speech recognition.At present,most of the researches using surface EMG signal for speech recognition have the problems that the number of electrodes used was small and the position selection was not scientific enough.It is necessary to analyze the influence of electrode position on recognition performance during speech recognition to achieve a better choice of electrode position and quantity.At the same time,because people's pronunciation habits vary greatly,whether the placement of the EMG electrodes should be fixed remains to be considered.Aiming at the problem of the small number of surface EMG electrodes and different positions in the existing research,this paper proposed to use 120 high-density surface EMG electrodes to cover all the muscles related to the pronunciation of the subject's face and neck,and to integrate the complete muscles of the pronunciation process.All electrical changes were collected.Although there may be redundancy in this part of the signal,the EMG change information during the pronunciation process can be completely preserved,which is helpful for subsequent channel screening.In this article,firstly 12 features were matched with three classifiers,and four evaluation criteria were proposed to analyze the recognition performance.The results showed that the performance of Linear Discriminant Analysis(LDA)and Support Vector Machine(SVM)in this data set were significantly better than K-nearest neighbor(KNN).The classification performance of LDA and SVM were similar,but the running time of SVM when matching all features were longer than that of LDA.Therefore,LDA was more suitable for speech recognition in this data set.At the same time,among the 12 features matched by LDA,Waveform Length(WFL)could achieve the highest classification accuracy,sensitivity and F1 score.When its running time were only slightly higher than other features by 1?2s,It could be concluded that WFL had achieved the best match with LDA.On this basis,this article analyzed the influence of electrode position on speech recognition.First,the s EMG changes of the symmetrical channels on the left and right sides of the face and neck were compared during the pronunciation process,and it was found that the EMG signal changes at the symmetrical positions had a high degree of consistency.Moreover,the classification accuracy of the EMG signals using channels from the left side and the right side were highly consistent,indicating that the channel in the symmetric position might contain the same EMG change information,and there was no need to repeat the collection.Finally,the Sequential forward selection(SFS)algorithm was used to sort and search the 120 channels.The results confirmed that a small number of electrodes could be used to achieve better recognition performance for different subjects.The specific number and location of channels varied from person to person,but it was valuable to select the location of channels from the physiological point of view.To sum up,the study of silent speech recognition based on high-density surface EMG signals in this paper provided a good refeence for the selection of features and classifiers during subsequent silent speech recognition,and at the same time it provided a good idea for the selection of electrode position and quantity.