Every snore has a fingerprint

Your ear doesn't measure sound in hertz

Ask most sleep trackers what happened last night and they'll give you one number: how loud, how often. But a decibel meter is a blunt instrument — two completely different sounds can hit the exact same peak volume. What actually distinguishes them is shape: which frequencies are loud, which are quiet, and how that pattern evolves over a fraction of a second. That shape is the acoustic equivalent of a fingerprint.

To capture it, MEDBUG doesn't just log a volume — it builds a mel spectrogram: a map of energy across frequency and time. The "mel" part matters. It's not a linear ruler of pitch; it's built to match how your own cochlea actually hears.

Snoring, breathing and airway noise all live mostly in that crowded, low-frequency region where human hearing — and the mel scale — pays the most attention. Building the model's input around that scale isn't a stylistic choice; it's aligning the sensor with the part of the spectrum that actually carries the signal.

The average shape of a snore, aligned to its loudest instant. 2,397 real snore events, stacked and averaged. Energy concentrates in two bands — roughly 200–500 Hz and 600–900 Hz — with almost nothing above 1 kHz. That two-lobe shape is the "fingerprint" a simple dB meter throws away. And like a speaking voice, every sleeper's fingerprint is unique: it depends on the airway, how much the upper airway collapses, and how hard the body is working to breathe in.

Most tools only listen for one sound

Ask a typical snore-tracking app what it heard, and "snore" is usually the only answer on offer — anything that doesn't fit gets ignored or lumped in as noise. That's a problem, because a night of disordered breathing is loud in more ways than one: gasps as the airway reopens, sighs from the nervous system resetting, coughs, throat-clearing — each with its own acoustic shape, each potentially telling a different part of the story.

Because these events are subtle and time-consuming to hand-score, most research has historically had to ignore them at scale — there simply hasn't been a practical way to label thousands of hours of audio. MEDBUG's classifier changes that arithmetic: it distinguishes over 10 distinct acoustic subtypes automatically, every night, with no one listening in.

Twelve nights, five different sounds. Each colour is a distinct acoustic class the model separates out automatically. Snorting (violet) dominates, but gasps, sighs, coughs and throat-clearing all show up as their own identifiable signature — not noise to be filtered out, but data.

Where in the night — and which night — is the real story

Even with the right acoustic fingerprint, a single "average loudness" per night still hides more than it reveals. Snoring isn't a steady drone — it swells and fades within an episode, and whole nights differ from each other depending on position, alcohol, congestion, or nothing identifiable at all. A meaningful picture needs both axes: within a night, and across many of them.

The same person, twelve different nights. Solid line = median peak loudness per snore episode; shaded band = the p5–p95 spread. Some nights climb toward a rough patch; others peak early, dip, then recover. Stitched together, this is what "comprehensive" actually looks like.

Stretched across 12 consecutive nights, this becomes less like a single test result and more like a rhythm — clusters of rough patches, quieter stretches, occasional one-off outliers. That's information a single lab study, by design, can never capture: it's built to summarize a night, not to reveal how one night differs from the next.

A summary score tells you something happened. A timeline tells you when, how often, and whether it's getting worse. — On the value of continuous acoustic monitoring