Crackling lung sounds are associated with many pulmonary diseases. Their occurrence reflects the quality and the severity of the disease. An automatic method for crackle detection is developed, based on analysing the spectral stationarity of the lung sound. The method is validated by studying the crackles of 20 adult patients; 10 with fibrosing alveolitis (FA) and 10 with bronchiectasis (BE). The number of crackles detected by the automatic method in inspiratory cycles is compared to the number of crackles counted from time-expanded waveforms by two expert observers. The total number of inspiratory cycles studied is 117 and that of crackles 1064. The method has a sensitivity of 89 per cent and a positive predictivity of 88 per cent for patients with FA, and 80 per cent and 83 per cent respectively, for patients with BE. The linear correlation coefficients between the numbers of crackles counted by the automatic method and by the observers is 0.86 (p less than 0.001) for the patients with FA and 0.93 (p less than 0.001) for the patients with BE. The values refer to whole inspiratory cycles. The new automatic method seems reliable enough for clinical and scientific purposes. It enables a rapid and objective analysis of large materials with crackling lung sounds.