Performance evaluation of an automatic speech recogniser incorporating a fast adaptive speech separation algorithm

This paper addresses the performance evaluation of a speaker-dependent automatic speech recogniser (ASR) that employs a speech separation algorithm as a front-end processor. The ASR software used is Dragon NaturallySpeaking (NS) Professional Version 6.1. The word recognition accuracy of NS is known to be very sensitive to background noise due to competing speakers, as well as ambient and environmental disturbances. In this work, a reduced complexity fast-converging adaptive decorrelation filter (ADF) is used to successfully reduce the interference from competing speakers. The recognition accuracy of NS for speech utterances before and after front-end separation was measured. A significant improvement has been observed with the proposed frontend processing.