Enhancement of speech using bark-scaled wavelet packet decomposition

In this paper, we propose a speech enhancement system, which integrates a bark-scaled wavelet packet decomposition (BS-WPD), a soft-decision gain modi(cid:12)cation and a \magnitude" decision-directed estimation technique. The BS-WPD provides an overcomplete auditory representation, having a higher frequency resolution than the critical band decomposition. Speech is estimated by Wiener (cid:12)ltering in the wavelet packet domain, mod-i(cid:12)ed by the signal presence probability. We introduce a \magnitude" decision-directed estimator for the variance of speech, which is closely related to the decision-directed estimator of Ephraim and Malah. This estimator achieves, in the established process, a better trade-o(cid:11) between noise reduction and signal distortion. The proposed enhancement algorithm is tested with various noise types, and compared to a conventional log-spectral amplitude estimator. We show that noise can be further suppressed, while preserving its natural structure and the intelligibility and quality of the speech components.