Probabilistic approach using joint clean and noisy i-vectors modeling for speaker recognition

Additive noise is one of the main challenges for automatic speaker recognition and several compensation techniques have been proposed to deal with this problem. In this paper, we present a new "data-driven" denoising technique operating in the i-vector space based on a joint modeling of clean and noisy i-vectors. The joint distribution is estimated using a large set of i-vectors pairs (clean i-vectors and their noisy versions generated artificially) then integrated in an MMSE estimator in the test phase to compute a "cleaned-up" version of noisy test i-vectors. We show that this algorithm achieves up to 80% of relative improvement in EER. We also present a version of the proposed algorithm that can be used to compensate multiple "unseen" noises. We test this technique on the recently published SITW database and show a significant gain compared to the baseline system performance.