DUAL-MICROPHONE ROBUST FRONT-END FOR ARM'S-LENGTH SPEECH RECOGNITION

This paper describes a novel method of improving the performance of a speech recognition front-end in non-stationary background noise. A two-microphone array has been designed that both enhances the speech and provides a continuous estimate of the background noise. This processing has been integrated with the standard ETSI DSR Advanced Front End so that the continuous noise estimate is an input to the first stage of Wiener filtering, increasing the noise suppression and hence the recognition performance. Tests with real-world noise have shown the recognition error rate is reduced by up to 50% when compared with single-microphone input to the same Advanced Front End.