A Noise Reduction Technique for Audio Signals Using Multi Stage Spectral Subtraction

This paper presents a technique for reducing additive noise from audio signals which have nonstationary properties. In the technique, we model a nonstationary signal as a set of tones having various durations. Since it is important to choose an appropriate analysis window size in Spectral Subtraction (SS) for such nonstationary signals, we propose a multi-stage SS technique, in which different window size is used in each Stage and the input signal is decomposed into stationary and nonstationary components according to their duration periods. Moreover, we also propose a technique for suppressing pre-echoes which appear in the SS with a long analysis window. To demonstrate effectiveness of the proposed approach, we implemented a 4-stage SS and evaluated its subjective performance for audio signals including speech and piano.