Spectral Subtraction of Nonstationary Audio Signals for Noise Reduction

摘要

This paper discusses a technique for reducing additive noise from audio signals which have nonstationary properties. Although spectral subtraction (SS) techniques are very effective for noise reduction of stationary signals, it does not work well when original signal has nonstationary properties such as pulsive wave sources and attack parts of piano sounds. We show that it is useful to choose an appropriate analysis window length in SS for such signals, and propose a multi-stage SS technique into which an window length adaptation mechanism is incorporated. We also propose a technique for suppressing pre-echo in SS with a long analysis window. A 4-stage SS is implemented and a subjective evaluation for 4 audio signals was performed. It was found that the multi-stage SS can produce higher quality audio signals than single-stage SS.