Bayesian noise compensation of time trajectories of spectral coefficients for robust speech recognition

Abstract

Our work presents a novel data driven compensation technique that modifies on-line the incoming spectral representation of degraded speech to approximate the features of high quality speech used to train a classifier. We apply the Bayesian inference framework to the degraded spectral coefficients based on modeling clean speech linear-spectrum with appropriate non-Gaussian distributions that allowmaximum a-posteriori (MAP) closed form solution to be set.MAP solution leads to a soft threshold function applied and adapted to the spectral characteristics and noise variance of each spectral band. We perform extensive evaluation of our algorithm against white and coloured Gaussian noise in the context of Automatic Speech Recognition (ASR), and demonstrate its robustness in adverse conditions. The enhancement process comes at little to no extra computational overhead, thus achieving real time, on line performance.