Spectral Subtraction is widely used for enhancing noise-corrupted speech signals. One of the main problems are "musical tones" in the filtered speech signal. Many solutions have been suggested to get rid of musical tones but they often lead to a poor speech quality or need cumbersome optimizations. We present a new approach based on nonlinear median filtering, which is easy to implement and especially efficient for suppressing musical tones without degrading the speech signal.