A non-linear digital filter system is proposed for automatic extraction of crackles which are discontinuous paroxysmal sounds in lung sounds. This system is composed of two filters: one is a stationary-non-stationary separating filter and the other is a width separating filter. The former separates non-stationary signals from stationary ones, using the prediction error for the input. This filter can roughly separate the crackles as non-stationary signals. The latter, the width-separating filter, performs signal extraction on the basis of the interval between the zero-crossing points of the non-stationary signals, simply realized by logical algebra. This filter extracts a small-width impulsive signal and its succeeding waves; such a waveform is typical of crackles. Crackles can be more precisely extracted from the non-stationary signals by this filter. Both of the filters are realized quite simply. The high performance of this system is shown in processing actual lung sound data.