Spectral noise subtraction with recursive gain curves

A well known technique for speech enhancement is spectral subtraction. This technique is attractive due to its simplicity and its low computational cost. However, spectral subtraction has the drawback of generating residual noise with a musical character, the so-called musical noise. In this paper, we propose a simple modi cation of the lter coe cient calculation in form of a recursion of the previous lter coe cient. As indicated by objective and subjective tests, the new recursive enhancement scheme substantially reduces musical noise without a noticeable impairment of the speech signal.