Speaker recognition in reverberant enclosures

This paper evaluates the effects of room reverberation on two automatic speaker verification (ASV) applications. Reverberation is simulated using an image method. AV is conducted with a multiple binary classifier model. Results are strongly dependent on speaker location, room size and reverberation time. ASV is poor given anechoic training but reverberant test speech. In voice activated security access, speaker locations can be identical for ASV training and testing, eliminating differences between respective impulse responses. Reverberation does not affect ASV even when it renders speech unintelligible. In covert speaker verification, speakers are uncooperative and mobile, thus test room responses unknown. By subjecting training speech to an impulse response corresponding to the centre of the room, ASV is degraded by 5.45 instead of 13.7 percent should anechoic training speech be used. Subjecting training speech to responses corresponding to two room locations does not further improve ASV performance.