Noise reduction in whisper speech based on the auditory masking model

This paper presents the issue of whispered speech enhancement. Based on multi-band spectral subtraction method where the introduced musical residual noise occurs, the proposed approach performs parametric subtraction according to the WSS (Whispered Sensitive Scale) method that is particular for whispered speech processing and auditory masking model. The algorithm is characterized by a tradeoff mechanism between the amount of the whispered speech distortion, noise reduction, and the level of musical residual noise, which are determined by appropriate adjusting the subtraction parameters. Compared with traditional subtractive-type algorithms, the proposed method results in a significant reduction of musical residual noise. Finally, objective and subjective evaluations are implemented illustrating the improvements over traditional subtractive-type algorithms.