A proposed control scheme for a vocal prosthesis

The authors report on research into the recognition of speech utilizing the myoelectric signals exhibited by the muscles that shape the vocal tract. Myoelectric signals from four muscle sites were digitized along with the acoustic waveform. Energy, average magnitude, and standard deviation of the biopotentials served as inputs to a maximum likelihood classifier. Recognition is close to three times a priori for a ten-word set, with a 70-80% chance of recognition within the first five choices of the maximum likelihood algorithm.<
>