Combating Reverberation in Speaker Verification

Abstract

This paper investigates reverberation in speaker verification. Reverberant test speech degrades the performance of speaker verification algorithms. The effect of training with speech originating from reverberant environments that are different from that of the test speech is investigated. It was found that with the room impulse responses used, that training with speech that is less reverberant than the test speech always improved performance. A method was proposed to determine given a set of reverberant training speech segments corrupted with different transfer functions, which transfer function corrupted the speech. A classification rate of 96.5% was achieved