Enhancement of musical signals degraded by background noise, using long-term behavior of the short-term spectral components

The issue addressed is the restoration of musical recordings degraded by background noise. The author presents a new technique based on the observation that the mere knowledge of the short-time spectrum of the noisy signals is not sufficient to enhance correctly the unknown musical signal. A procedure that allows the detection of sinusoidal signal components too low to be detected on the short-time power spectrum is proposed. When a component is detected, the performance of the enhancement system can be improved by using a longer observation-time to estimate the corresponding frequency bin. This method makes it possible to estimate sound partials that would otherwise get distorted by the restoration process.<>