Use of myoelectric signals to recognize speech

It has been determined that the myoelectric signals (MES) from muscles associated with speech that are obtained using surface electrodes can be used to recognize speech at approximately five times a priori. The results of ten experiments have been evaluated using a maximum-likelihood recognition scheme and have consistently yielded similar recognition accuracy among several subjects. Individual word recognition from a ten-word set has approached 60%. Parameters studied have included energy, magnitude, and trial deviation. The results suggest that accuracies at a level suitable for use in a vocal prosthesis will require the development of a hybrid recognition algorithm utilizing either statistical and heuristic word separation or neural nets