Facial SEMG for Speech Recognition Inter-Subject Variation

The aim of this project is to identify speech using the facial muscle activity and without audio signals. The paper presents an effective technique that measures the relative muscle activity of the articulatory muscles. The paper has also tested the performance of this system for inter subject variation. Three English vowels were used as recognition variables. This paper reports using moving root mean square (RMS) of surface electromyogram (SEMG) of four facial muscles to segment the signal and identify the start and end of the utterance. The RMS of the signal between the start and end markers was integrated and normalised. This represented the relative muscle activity, and the relative muscle activities of the four muscles were classified using back propagation neural network to identify the speech. The results show that this technique gives high recognition rate when used for each of the subjects. The results also indicate that the system accuracy drops when the network trained with one subject is tested with another subject. This suggests that there is a large inter-subject variation in the speaking style for similar sounds. The experiments also show that the system is easy to train for a new user. It is suggested that such a system is suitable for simple commands for human computer interface when it is trained for the user.