Snoring is the hallmark of the obstructive sleep apnoea syndrome and several studies explore possible correlations between them. In this work an improved methodology with respect to [4] is proposed, based on a proper energy threshold applied on audio recordings for sound/silence detection, and on a feature vector of 14 elements (13 mel frequency cepstral coefficient plus the number of zero crossings) for sound classification. This feature vector is obtained from a 62-elements one by applying a genetic algorithm, fitted to obtain the best classification of the training/validation sets. The feature vector is analyzed by means of a radial basis neural network to perform snore events identification. Finally, formant frequencies and time analysis are also investigated to split up post-apnoeic snores and normal ones. Audio data from 26 patients of different age and sex are used to test the methodology: 6 patients (3 male and 3 female) were used to train the nets (1800 snores) and 4 patients to validate the classification (600 snores). On the whole dataset of patients, a sensitivity between 69% and 84% is obtained in the detection of post-apnoeic snores.

Index Terms. snore, neural network, Mel frequency cepstral coefficients, genetic algorithm, obstructive sleep apnoea