2012 Volume 50 Issue 1 Pages 38-46
This paper proposes the use of electrode grid for Japanese vowel recognition based on surface electromyography (sEMG). Previous studies have indicated the potential effectiveness of sEMG-based speech recognition, not only for healthy people, but also for dysarthric patients. In these studies, however, disc electrodes or parallel bar electrodes were used and located empirically, although there exist relatively small muscles in proximity to each other in the face or neck region. In order to avoid missing out information about speech, we examined the effectiveness of using an electrode grid, which consists of densely-spaced multielectrodes. In our experiments, we measured sEMG signals from the submental region with the electrode grid during the production of 5 vowel sounds. Continuous hidden Markov models were applied to the sEMG signals for vowel recognition. We compared the recognition accuracies between the two methods: One was based on signals from all channels and the other was based on virtually reconstructed single bipolar signal. The former achieved considerably higher recognition accuracy than the latter. This result indicates that using electrode grid is more effective in extracting information for sEMG-based speech recognition.