MMSE Feature Reconstruction Based on an Occlusion Model for Robust ASR

This paper proposes a novel compensation technique developed in the log-spectral domain. Our proposal consists in a minimum mean square error (MMSE) estimator derived from an occlusion model [1]. According to this model, the effect of noise over speech is simplified to a binary masking, so that the noise is completely masked by the speech when the speech power dominates and the other way round when the noise is dominant. As for many MMSE-based techniques, a statistical model of clean speech is required. A Gaussian mixture model is employed here. The resulting technique has clear similarities with missing-data imputation techniques although, unlike these ones, an explicit model of noise is employed by our proposal. The experimental results show the superiority of our MMSE estimator with respect to missing-data imputation with both binary and soft masks.