Geometric contamination for GMM/UBM speaker verification in reverberant environments

Reverberation generated by multi-path acoustic propagation in enclosures is one of the most critical issues for distant-speech speaker verification systems. While late arrivals can be treated as additive noise, early reflections critically affects the speech spectral properties that allow differentiating among speakers. Considering a standard GMM/UBM speaker verification system based on MFCC, a geometric data contamination scheme relying on a rough modeling of the sound propagation through the image method is presented to attack the reverberation. The contribution of this paper is twofold: first, an analysis of the amount of knowledge of the impulse response needed for an effective contamination is carried out on simulated data; second, the geometric contamination is applied to real impulse responses to validate its use in real application contexts. Preliminary experimental results are provided to support the theoretical study and show the effectiveness of the approach.