This paper describes the automatic recognition method which classifies the respiratory sounds into seven categories : vesicular, bronchial, sonorous, sibilant, coarse rale, fine rale and medium rale. This method enables automatic auscultation which will improve mass survey and screening tests of the respiratory diseases. In order to analyze sound data, which were obtained from patients with an improved electret microphone, three calculation processes were adopted and examined : FFT, auto-regressive linear prediction (AR), moving average linear prediction (MA), where sampling time was 200 or 250 μsec.
The tests showed the MA process to be the best, and applicable even to mixed respiratory sounds. However as this calculation was often too difficult to converge into a clear solution, an approximate MA process was developed to render calculations easy enough for practical use.
Multivariate normal density functions were used to classify the data because each typical respiratory sound showed almost normal distribution. About 30 samples of isolated respiratory sounds were correctly classified into the seven distinct categories. Except for the following combinations, more than 40 samples of the artificially mixed respiratory sounds using R. Murphy's medical training tape were separated and classified : (a) sonorous and rale (medium, coarse, fine), and (b) sibilant and fine rale.
Conclusion : this method is applicable of mass survey and screening tests of respiratory diseases.
