The Rumble We Never Explained

Everyone knows cats purr. Almost no one can tell you how, and for good reason: the mechanism was quietly reopened in 2023, and the why was never as simple as "the cat is happy." Below is the purr band you can hear, twenty real recordings measured from scratch, the cry hidden inside them, and a plain accounting of what we actually know.

A purr is one of the strangest sounds an animal makes. It is nearly continuous, running on the out-breath and the in-breath with barely a seam, for minutes at a time. Its pitch is absurdly low and absurdly stable: a fundamental down around 25 hertz, at the very floor of what a human ear will register as a tone.¹ And the animal producing it weighs four kilograms. Measured here from twenty openly licensed recordings, that pitch comes out at a median of 27.2 Hz, with every recording the instrument could read landing inside 20 to 30 Hz.⁷ So: how?

The mechanism nobody had pinned down

For half a century the standard answer was neural. A cluster of neurons in the cat's nervous system acts as a central pattern generator — a biological metronome — sending bursts of signal to the muscles of the larynx (voice box) roughly 25–30 times a second. (The exact seat of that oscillator was inferred from the muscle recordings, never pinned to a single anatomical spot.) Each burst snaps the glottis partly shut; air passing through is chopped into a train of little puffs, and that train is the purr. On this picture the cat is, in effect, drumming its own vocal folds with its nervous system. This is the model built up across the classic work of the 1970s–90s.¹

It had one nagging problem. Producing a 25 Hz tone by neural drive means firing a muscle cleanly, on schedule, 25 times a second, for minutes — an unusually fast, tireless bit of motor control for a mammal. It was assumed, not shown, that the brain was doing all of it.

In 2023 a team led by Christian T. Herbst took eight larynges from domestic cats (euthanised for unrelated medical reasons) and blew warm, humid air through them on a bench — no brain, no nerves, no muscle activation at all. The isolated larynges purred: they self-oscillated at 25–30 Hz, squarely in the natural range. The vocal folds turned out to carry unusual masses of connective tissue — "pads" — that let them flap at very low frequency on airflow alone, the same myoelastic-aerodynamic principle behind human "vocal fry." ²

Here is the honest shape of it, and the part most headlines got wrong: the 2023 study did not prove that a live cat purrs without any muscle involvement. Excised larynges show what the tissue can do unaided — that a neural metronome is not necessary to reach purr frequencies. In the living animal, muscle tone almost certainly still sets up and sustains the oscillation (the folds have to be held in the right posture for the airflow to grab them). What flipped is the burden: the ~25 Hz rhythm no longer needs a dedicated neural clock ticking every cycle. The authors say so themselves: their data, in their words, do not fully reject the active-muscle model, and the purr is potentially augmented by muscle contraction.² Whether the live purr is mostly passive self-oscillation, mostly active gating, or a blend that shifts moment to moment is still open. This is a live disagreement, freshly reopened — not a solved problem you were simply never taught.

Why they purr — also not what you think

The comfortable story is that purring means contentment. It often does. But cats also purr when they are injured, frightened, giving birth, being examined by a vet, and dying. Whatever purring is for, "I am happy" cannot be the whole account — a signal that shows up at both the best and worst moments of a life is doing something more general. The leading reading is that a purr is a request and a self-soothing behaviour: a low-cost signal that says stay with me, attend to me, produced in states where the cat wants exactly that.³

The sharpest evidence is the solicitation purr. In 2009, Karen McComb and colleagues studied the specific purr cats use to get fed, the one owners describe as impossible to ignore. Hidden inside its low rumble is a high-frequency peak, around 220–520 Hz with a pooled mean of 380 Hz, uncannily close to the frequency band of a human infant's cry. People rated purrs carrying that peak as more urgent and less pleasant, and when the peak was digitally removed and the purr replayed, the same listeners rated it less urgent.⁴ That is the peak you add with the reveal the cry slider above.

Two corrections to how that study is usually retold, both taken from its own supplementary table. First, the peak is not something a hungry cat adds and an ordinary one lacks: its frequency did not differ significantly between solicitation and non-solicitation purrs (391.70 Hz against 367.50 Hz, P = 0.284). What differed was its height, 3.59 dB against 0.51 dB (p = 0.008). The cat is not smuggling in a new sound; it is turning up one that was already there. Measured below on twenty recordings, none of them labelled as solicitations, the peak is duly present in ordinary purring too. Second, the playbacks were peak-normalised to 99% of full scale, which equalises amplitude rather than perceived loudness; "matched loudness" overstates it.

What twenty recordings actually say

Everything above this line is quoted. The instrument at the top of the page plays numbers taken out of four papers, and until 2026-08-24 this page had never listened to a cat. So here is the other kind of thing: twenty openly licensed recordings of domestic cats purring, measured from scratch, with the apparatus in the open and every recording reported whatever it said.⁷

The order matters more than any single number. First the confidence thresholds were fixed, by running the estimators over three thousand stretches of noise containing no periodicity at all (flat, 1/f, and randomly spaced bursts) and taking the level none of them reached. Nothing in that step has heard a cat. Then the estimators were tested on synthetic signals whose answer was known, including the ones built to fool them. Then, before any of the twenty were opened, the whole thing was graded against an answer key made by somebody else.

Nineteen of the twenty produced a fundamental, and all nineteen land inside the 20 to 30 Hz the literature claims: median 27.15 Hz, range 23.98 to 28.87. Frazer Sissom, Rice and Peters put domestic cats at 26.3 ± 1.95 Hz in 1991.¹ The twentieth is reported as no answer rather than as a gap, which turns out to be right: it is also the only recording whose reading moves when you re-encode it.

Graded against someone else's answers

Herbst and colleagues published, alongside the 2023 paper, a 12.4 second recording of a live domestic cat purring (13 years old, castrated male) together with their own annotation of it: an interval tier separating the ingressive and egressive phases of each breath, and a point tier marking 157 individual purr cycles by hand. As far as this study could find, it is the only openly licensed, cycle-annotated purr recording in existence, and it is the closest thing to a calibration standard the field has.²

Run blind on that recording, the cycle counter put 128 of their 157 marks within 5 milliseconds, with a mean offset of 2.05 ms and a scatter of 0.91 ms, and read the rate as 24.881 Hz against the 25.006 Hz their marks give. That is 0.5% apart. Select the first recording in the picker above and the two sets of ticks are drawn together: green for the counter, white for theirs.

That 5 ms is not a flattering choice, it is the only honest one, and finding that out was the most useful thing that happened here. The first version of this check used a 20 ms tolerance and reported 144 of 157, which is 91.7% and sounds excellent. Then a deliberately adversarial control, one of a set written to corrupt each input and confirm the checks go red, was pointed at the same comparison with the answer key replaced by 157 marks scattered at random. It refused to go red. Of course it did: the detections sit about 43 milliseconds apart, so at a tolerance of 20 ms almost every instant in the recording is within reach of one, and a random key scores 90.4%. The observed 91.7% was a lift of 1.01 over chance, p = 0.35. It was not evidence of anything.

Computed properly, against 2,000 random keys at four tolerances, the picture is both worse and far more convincing: at 5 ms the counter finds 81.5% where chance finds 23.0%, a lift of 3.5 with p = 5 × 10⁻⁴, and at 2 ms it finds 32.5% where chance finds 9.2%. The claim that survives is the smaller one, and it is the one worth having. The rate, meanwhile, is pinned much harder than the count: across the sweep's plateau the recall moves from 81.5% to 89.2% while the rate moves by a third of a percent, which says the residual is two marking conventions disagreeing about what counts as one event rather than an instrument reading the wrong speed.

A 1991 result, recovered from 2023 data

The abstract of How cats purr contains a sentence that has been sitting behind a paywall since 1991: "The frequency at mid-expiration exceeds that at mid-inspiration by 2·4 ± 1·3 Hz."¹ A purr, in other words, is slightly faster on the way out than on the way in. Herbst's labels make that testable on their recording, and it holds: computed from their marks and their phase labels, the egressive cycles run at 26.184 Hz (n = 50) against 24.601 Hz ingressive (n = 97), a difference of +1.58 Hz, permutation p = 5 × 10⁻⁵. That is within one standard deviation of a number published thirty-five years earlier, on a different cat, by a different group, in a paper the present study could not even read past its abstract.

And then it does not generalise. Split all twenty recordings blind, by loudness rather than by any label, and the louder half of the breath is the faster half in 11 of 19. The difference reaches p < 0.05 in eight of them: five going up and three going down. The median difference across recordings is 0.08 Hz, which is nothing at all. The effect is large and real inside an individual cat, and has no consistent direction across cats. Which is exactly what the literature does when you read it instead of summarising it: Remmers and Gautier report the opposite direction to Frazer Sissom, and Schötz and Eklund found both directions among four cats.⁸

The cry is in the ordinary purrs too

If McComb's own numbers are right that the voiced peak differs between contexts in height rather than in frequency, it should be findable in purrs nobody has labelled. Nothing in this corpus is labelled. It is there: the 220 to 520 Hz band carries a median 4.15 dB more energy than the spectrum around it predicts, against 0.70 dB for a matched control band at 600 to 900 Hz, and it is the higher of the two in 16 of 20 recordings (sign test p = 0.012). Of the sixteen whose excess peaks inside the band rather than on its edge, the peak sits at a median of 399 Hz. McComb's pooled mean is 380.

One objection deserved a measurement rather than a caveat. Eight of the twenty are MP3 or Vorbis, and psychoacoustic codecs are least careful in exactly the bottom octave where a purr lives. So the lossless recordings were pushed through the same encoders the lossy ones arrived in and re-measured: same cat, same second, only the encoder changed. Across 25 comparisons the median shift in the fundamental is 0.014 Hz. The single recording it destabilises is the one the confidence floor had already refused.

Last, the thing this is worth. The published measurements of domestic-cat purr frequency that state their own sample size are remarkably few: four cats in Schötz and Eklund, one in Eklund, Peters and Duthie, one annotated in-vivo recording in Herbst.⁸ Frazer Sissom gives 26.3 ± 1.95 Hz with no recoverable n. The largest modern corpus, 557 purrs from 21 cats, measured duration and spectral centroid and did not measure the fundamental at all.⁹ The number everybody quotes rests on a small handful of animals. These twenty are opportunistic internet recordings of unknown provenance, no substitute for a designed study, and possibly not even twenty distinct cats. They are also twenty more measurements than existed in the open yesterday, with the code, the sources and the failures attached.

The claim to be careful with

You have probably read that purring heals — that the 25–150 Hz vibration mends bone and tissue, which is why cats recover so well. It is a lovely idea and it is not an established fact. The specific "purr as healing mechanism" claim traces to a single 2001 conference presentation by Elizabeth von Muggenthaler, not a controlled, peer-reviewed study showing that purring heals cats.⁵ Separately, there is real bone-biology research showing that certain low-magnitude, high-frequency mechanical vibrations can influence bone in the lab.⁶ But that work is not about cats, not about purring, and does not demonstrate that a cat's own purr heals its own body. The frequencies overlap; the evidence does not connect. Treat "purring heals" as an appealing hypothesis — the honest label — not a finding.