7th International Conference on Spoken Language Processing

September 16-20, 2002
Denver, Colorado, USA

Speech Enhancement Based on Combining Perceptual Enhancement and Short-Time Spectral Attenuation

I. Potamitis, Nikos Fakotakis, George Kokkinakis

University of Patras, Greece

The present paper introduces a cooperating framework between Perceptual Enhancement (PE) and Minimum Mean Square Error Log-Spectral Amplitude Estimation (MMSE-LSA). The link between the two aforementioned spectral modification algorithms is a Speech Absence Probability (SAP) responsible for weighting the fusion of both spectral gain modification algorithms for each spectral bin. It has been noted that PE eliminates residual noise at the expense of spectral deformation of unvoiced speech and high-frequency speech components. On the other hand the LSA technique is based on modeling assumptions that lead to smaller degradation than PE at the expense of some residual noise and an echo-like processing artifact. The proposed hybrid scheme leads to a reduction of residual noise component compared with LSA and a reduction of spectral distortion observed with PE when it is employed at low SNRs. We evaluate the proposed algorithm using White Gaussian and car-noise on the task of improving the quality of speech and demonstrate its robustness at low SNRs.

Implementation code and demo files are available at: http://slt.wcl.- ee.upatras.gr/Potamitis/web/LSA_PE_index.htm


Full Paper

Bibliographic reference.  Potamitis, I. / Fakotakis, Nikos / Kokkinakis, George (2002): "Speech enhancement based on combining perceptual enhancement and short-time spectral attenuation", In ICSLP-2002, 1785-1788.