EUROSPEECH 2001 Scandinavia
This paper presents a novel data driven compensation technique that modifies on-line the incoming spectral representation of degraded speech in order 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 the modeling of clean speech linear-spectrum with appropriate non-Gaussian distributions that allow maximum a-posteriori (MAP) closed form solution. The MAP solution leads to spectral magnitude estimation adapted to the spectral characteristics and noise variance of each spectral band. We perform extensive evaluation of our algorithm using white and coloured Gaussian noise on the task of improving the quality of speech perception as well as Automatic Speech Recognition (ASR), and demonstrate its robustness at very low SNRs. The enhancement process comes at little to no extra computational overhead for ASR systems, thus achieving real time performance.
Bibliographic reference. Potamitis, I. / Fakotakis, Nikos / Kokkinakis, George (2001): "Map estimation for on-line noise compensation of time trajectories of spectral coefficients", In EUROSPEECH-2001, 1899-1902.