If you want to use technology to identify when a person is snoring, you’d probably need a long series of steps that begins with attaching wires to the person’s scalp, chin and eyelids. But soon the task will be much simpler — just stick a microphone on the bedside table, and use a computer to distinguish what’s a snore sound and what’s not.
Credit for this breakthrough goes to William Duckitt, Seppo Tuomi and Thomas Niesler of the University of Stellenbosch. They describe their technical tour de force in a paper called Automatic Detection, Segmentation and Assessment of Snoring from Ambient Acoustic Data, published in the journal Physiological Measurement.
The traditional approach, called polysomnography, electrically measures eye movements, heartbeats, the airflow of breathing and other physiological activities. It’s complex and cumbersome.
Duckitt, Tuomi and Niesler go at it differently, using mathematical “methods that have proved effective in the field of speech recognition”. After a bit of tweaking, these identify the snores, if there are any, within an audio recording.
The scientists recorded two slumbering women and four slumbering men. Altogether they obtained nine hours of “audio containing frequent snoring”.
The goal was to (a) have the computer analyse all that sound and then (b) compare the machine’s performance with that of a well-trained human being.
The human snore analyst listened to all those recordings and, moment by moment, indicated what was a snore and what wasn’t. This person had to remain consistently alert while monitoring all those hours of recorded breathing, snoring and duvet noises.
“Duvet noises” is a phrase coined by Duckett, Tuomi and Niesler. They define it as “the sounds made by the bed linen when the subject moves during sleep”.
The scientists divided the sounds of the night into five types: duvet noise, snoring, breathing, silence and “other noise”. They hope to eventually refine their techniques so a computer will reliably identify the most common kinds of “other” sounds — car noises, barking dogs and, especially, sleep talking.
For now, the basics are enough of a challenge. “It is sometimes very difficult to decide where loud breathing ends and snoring starts,” the study laments. “Periods of silence are most often misclassified by the system as breaths or duvet noises. Similarly, duvet noises are often misclassified.”
Subtleties abound. This, for instance: “A snore may fade to become a breath. Hence, a snore is not always followed by a silence. A breath may develop into a snore. Hence, a snore is not always preceded by a silence.”
The human snore analyst identified 5 560 instances of snoring, 4 190 non-snoring breath sounds, 133 duvet sounds, 9 421 periods of silence and 221 “other noises”.
The computer concurred with 82% to 89% accuracy, say the scientists.
This raises hopes that, with a little more refinement, the system will accurately report that most intriguing of snoring statistics, the number of snores per hour, the delightfully named “snoring index”. Not long ago, such capability might have seemed just a silly dream.
Marc Abrahams is editor of the Annals of Improbable Research and organiser of the Ig Nobel prize