Voice command recognition system for handicapped persons using simple multi-layered networks

A portable voice command recognition system for a specific speaker using a multilayered network has been developed to reliably recognize up to 16 commands. The intended application is the control of various program modes of a multichannel muscle stimulator. The voice input system consists of microphone and a detector which produces a pulse for each zero-crossing of the filtered microphone voltage. A microprocessor transforms these pulses into an array of 32 zero-crossing frequencies. These data were used for conventional pattern matching as well as for several network methods of word recognition. The recognition accuracy was tested by pronouncing each of 16 command words 50 times and measuring how often the computer responded with the correct answer. The pattern-matching method achieved an accuracy of 88%. The network method, with two layers and 15 repetitions during the training period, achieved an accuracy of 92%. A four-layered network produced an accuracy of 97%.<>