Bit stream based wireless speech recognition using mixed excitation linear prediction (MELP) vocoder

Abstract

Wireless speech recognition techniques have been the focal point of considerable research activity in the last decade. Most of the previous research activity assumes that the communication channel is only large enough to transmit either speech coding or speech recognition parameters. A bit-stream based front-end for a wireless speech recognition system that operates on a mixed excitation linear prediction (MELP) vocoder is presented. Speaker dependent, isolated word recognition accuracies obtained from conventional and bit-stream based front-end systems are obtained and their statistical significance is discussed. The recognition accuracies prove that the bit-stream based front end gives comparable performance to that obtained with the original input speech and is superior to the recognizer obtained from low bit rate decoded speech. The proposed system's performance in wireless channel conditions is also evaluated, via simulations, in order to observe adverse channel effects; these simulation results prove that the bitstream based front-end is useful for wireless systems.