Adaptive filtering for attenuating musical noise caused by spectral subtraction

A method of alleviating processing distortion caused by spectral subtraction is presented. It is well known that the spectral subtraction introduces annoying artifacts, which are referred to as undesirable musical noise, in the enhanced speech. The enhancement quality of the spectral subtraction quite depends on the performance of reducing the musical noise. Our approach exploits an adaptive filter in order to eliminate such distortion. In the method, the enhanced speech obtained by the spectral subtraction is used as a reference signal of the adaptive filter. The proposed method utilizes the characteristic difference between the musical noise and speech, i.e., the majority of the frequency components consisting the musical noise have the duration shorter than those of speech. Therefore, when the convergence speed of the adaptive filter is slower than the lifetime of the musical noise but faster than that of speech, only speech components can be tracked by the filter while the musical noise components are attenuated. Simulation results show that the proposed method can efficiently reduce the musical noise and the enhancement quality is improved in comparison with the conventional spectral subtraction.