A single channel speech enhancement technique using psychoacoustic principles

This paper presents an algorithm for robust noise reduction technique and speech enhancement based on psychoacoustic principles for low SNR corrupted speech signals. Assuming the clean speech components as masker for the distortions in the enhanced speech, a good quality of the enhanced speech can be obtained, if the distortions in the enhanced speech are kept below the masking curve of the clean speech. But for a single channel noise reduction technique, the real clean speech is not available. The speech components required to compute the desired masking curve are therefore estimated from the corrupted speech using advanced spectral subtraction techniques. The computed masking curve, the estimated corrupting noise and a time-frequency dependent parameter to control the desired predefined residual noise are then used to derive the weighting function. The proposed approach performs well for low SNRs of the corrupted speech compared to other approaches. This is confirmed with the subjective listening and objective grading tests. Some audio demonstrations are available from: http://www.hsu-hh.de/ant/index_74BUg40IeCJD04Zr.html.