Towards a home apnea monitor: a new method of extracting speech and voiced-snore segments from noisy respiratory sounds

Snoring is the most frequent symptom of obstructive sleep apnea (OSA), which is a difficult-to-diagnose serious disease of high prevalence. Recent work from our group (and others) has unequivocally established that snore-related-sounds (SRS) carries sufficient information to base the design of a community screening device on SRS. A main challenge to overcome in such a device is that the background electrical and acoustical noise, including embedded speech segments, can corrupt genuine snores. In this paper, we model snore sounds as the response of a mixed-phase system, when excited by a pseudo-periodic or white noise sequences (as needed to describe voiced and unvoiced snore segments). Observation noise is considered a Gaussian process. We present a novel algorithm to extract speech and voiced-snore segments from corrupted SRS measurements. The algorithm utilizes higher order statistics (HOS) due to its insensitivity to Gaussian noise and the ability to reconstruct a system preserving true phase characteristics.