A Comparison Study of Spectral Subtraction Speech Enhancement Methods

Spectral subtraction is a widely employed method for the suppression of additive noise in speech signals. Several variants of the basic approach have been proposed over the years to address certain shortcomings, chiefly the quality of the remnant noise and its trade-off with speech distortion. In this paper, we present a unified view of the various forms of the spectral subtraction speech enhancement method. Typical realizations chosen from each of three broad categories are described. Results of objective and subjective testing of performance are reported and discussed. We show that mild over-subtraction of noise, followed by spectral weighting, can achieve a better speech quality-intelligibility trade-off than a purely over-subtraction-based approach.