Obstructive sleep apnea (OSA) is a sleep disorder associated with several anatomical abnormalities of the upper airway. Our hypothesis is that it is possible to distinguish between OSA and non-OSA subjects by analyzing particular speech signal properties using an automatic computerized system. The database for this research was constructed from 90 male subjects who were recorded reading a one-minute speech protocol immediately prior to a full polysomnography study; specific phonemes were isolated using closed group phoneme identification; seven independent Gaussian mixture models (GMM)-based classifiers were implemented for the task of OSA / non-OSA classification; a fusion process was designed to combine the scores of these classifiers and a validation procedure took place in order to examine the system’s performance. Results of 91.66% specificity and 91.66% sensitivity were achieved using a leave one out procedure when the data was manually segmented. The system performances were somewhat decreased when the automatic segmentation was used, resulting in 83.33% specificity and 81.25% sensitivity.

Index Terms. Obstructive sleep apnea, speech signal processing, speaker recognition, phoneme identification.