Wideband speech enhancement approaches using a Kalman filter and a perceptual post-filter

Abstract

This thesis discusses the issue of speech enhancement in the presence of white noise for wideband speech. Speech enhancement aims at improving one or more perceptual aspects of speech such as overall quality, intelligibility, or listener fatigue. For example, reducing the amount of background noise in an automatic recognition system might improve the speech recognition rate by a considerable amount improving the intelligibility for machine recognizers. Most speech enhancement algorithms have been developed and tested for conventional narrowband speech systems (sampling rate of 8 kHz). This thesis shows that some narrowband algorithms can be extended to be used for noise reduction in wideband speech applications (sampling rate of 16 kHz) by presenting a slightly modified Kalman filter based approach that uses a perceptual post-filter. The results are compared with the performance of other speech enhancement algorithms. The results show that the proposed method has a good potential for speech enhancement in wideband, especially in lower SNRs environments.