Improving Speaker Verification Robustness by Front-End Diversity and Score Level Fusion

Abstract

In this paper, we studied the impact of the mismatch existing between training and testing data due to the presence of an additive noise on the performance of speaker verification system. Using a GMM-UBM system with MAP adaptation as a baseline system, front-end diversity is achieved by using MFCCs and different asymmetric MFCCs stand-alone as features or followed by PCA and LDA as dimensionality reduction techniques applied before the GMM-UBM back-end classifier. A score level fusion framework based on logistic regression is proposed to improve performance and to mitigate noise degradation. The obtained results on both clean and corrupted TIMIT database confirm the superiority of fused system in clean and noisy environment against each system alone, and the drastic degradation of the performances of PCA and LDA based systems in the presence of environmental noise.

A novel speaker verification approach for certain noisy environment

“…Improvements were based on new construction of filter bank according to human factor. In addition to comparison of different front-ends, N. Asbai, put forward a score level fusion framework based on logistic regression to improve performance and to mitigate noise degradation [4]. As to score-based normalization, A. Nautsch introduced a cohort normalization technique using MFCCs and Hidden Markov Models (HMMs) [5].…”