Robust speaker recognition by means of acoustic transmission channel matching: An acoustic parameter estimation approach

Mismatched acoustic transmission channels are known to cause degradations in reliability of automated speaker recognition. The problem becomes significant in the presence of non-trivial ambient noises and acoustic reverberation. Channel equalization, i.e. the removal or reduction of the channel effects, to some extent, mitigates the mismatching problem at the cost of added distortions to the vulnerable speech signals themselves, and therefore its effectiveness is limited. Conversely, this paper proposes to estimate noise and reverberation and incorporate them into individual training examples to create virtually matched channels. A "fine tune" training procedure is performed before final decision making. The paper details the proposed method, presents the results and discusses the potentials and limitations.