A method for impact noise reduction from speech using a stationary-nonstationary separating filter

A method for reducing impact noise mixed into speech is proposed. This method first detects noisy part of the input signal, contaminated with impact noise, using a nonlinear digital filter named as a stationary-nonstationary separating filter, and then applies time-frequency domain masking only to the noisy parts. The time-frequency domain masking is realized with a voice model and a noise model. The voice model is generated from both of training speech data and the part of the input signal, judged as a clean part where noise is not involved. The noise model is generated from training noise data. These two models are utilized to determine the masking function. Computer simulations verify the high performance of the proposed method.