Multi-stage spectral subtraction for enhancement of audio signals

This paper describes a technique based on spectral subtraction (SS) for enhancing audio signals which have nonstationary properties. In the technique, audio signals are modeled as a sum of sinusoidal components, which are called tones, of various frequencies and durations. We show that the analysis window length for SS should be adjusted in accordance with the tone duration to obtain higher performance on SS. Then we propose a multi-stage SS approach into which an appropriate window length selection mechanism is incorporated. We also propose a technique for suppressing an artifact called pre-echo which appears in the noise reduction process. It is shown from the result of a subjective evaluation test that the proposed technique gives superior performance to the conventional SS for enhancement of audio signals including speech, piano, drums, and orchestra.