SCIENCE / SUMMARY / ACOUSTIC SNORING

How accurately can a phone recording at home count snoring?

SINGLE STUDY · ALGORITHM VALIDATION ■■□□ low evidence: one centre, 55 people in the test set and one phone model

Who

214 people at one university hospital in Korea between January 2019 and February 2023. 105 were recorded in hospital (phone: LG G3) and 109 at home (phone: iPhone 11); in both settings the phone was placed 1 metre from the person. The 55 people in the test group had a mean age of 46 and a mean AHI of 13.2 per hour; 35 of them had an AHI below 15.

What they did

For each person, 200 minutes of audio scored as sleep by the sleep test (400 slices of 30 seconds) were taken; slices spent awake were left out. Two trained listeners each heard every slice separately, and a slice counted as "snoring" only if both heard snoring. The model was trained on the hospital recordings plus home recordings from 54 people and tested only on home recordings from the remaining 55; no person appeared in both training and testing.

What they found

Per slice, sensitivity was 89.8% and specificity 91.3%. The share of the night spent snoring, observed versus predicted, was strongly related (r = 0.97; 95% CI 0.95 to 0.99), with a mean absolute error of 5.09 percentage points. The mean observed snoring share was 24.5% and the predicted share 28.6%. In the two subgroups snoring more than half the time (5 people each), specificity fell to 0.72 to 0.78. Among people with an AHI of 15 or more the correlation was r = 0.94.

Limits

As the authors note: there is no universally accepted definition of snoring and the result rests on human labels; the heavy snoring subgroups are small; participants were mostly middle aged and overweight; the phone specific setup may limit how well results carry over to other devices and homes. In addition: since awake slices were excluded, how the model handles sounds made while awake was not tested; one subgroup's specificity appears as 0.942 in the table and 0.952 in the text. Funding: a hospital research grant. No conflicts of interest declared; four authors are affiliated with the Asleep Research Institute.

Link to Nefesly

It shaped how we show snoring time: labelling based on moments where two people agree, and testing the model on home recordings of people it did not see in training.

Reference

Hong J, Yang SK, Kim S, et al. Real-Time Snoring Detection Using Deep Learning: A Home-Based Smartphone Approach for Sleep Monitoring. Nature and Science of Sleep. 2025;17:519-530. doi:10.2147/NSS.S514631 · Open access (CC BY-NC 4.0).

Why chosen

One of 15 sources selected from 534 candidates on the acoustic snoring and apnea axis; a study testing phone snoring detection on home recordings with a split by person.

Summarised by

Nefesly, from the full text, 11 October 2026. Awaiting clinician check.

Copyright

Summarised from the full text, not translated from the abstract.