0

Four seconds of quiet, then a dozen real hits. Afterwards it asks how many you actually hit — correcting that number is what turns a guess into a measurement, because it says which sounds in that recording were yours and which were the beach.

It changes a setting only where your paddle demonstrably differs from the usual one. On a blade that rings almost three times as long, calibrating took the count from 10 hits in a minute to 43 of the 55 actually played. On a paddle that already matches, it correctly changes nothing.

Not the settings. How far the hits stand over the background — the headroom — decides almost everything, and it is set by where the phone is, not by anything in this app.

To measure it, 123 real hits were cut out of the recordings, cleaned of their own background, and replanted into 36 built beaches at a controlled headroom from 10 dB to 36 dB — 2,520 hits in all, with the exact time of every one known. The hit is real, the answer is exact, and the only thing that changes down the table is how loud the beach is.

Read it as a cliff, not a slope. Below about 15 dB it counts almost nothing; every decibel between 15 and 25 buys a great deal; past 30 it is at its ceiling and the last row is slightly worse, because in a close, quiet recording other sharp sounds clear the same tests. Four decibels of headroom is roughly halving your distance to the court, or getting the microphone out of the wind. No slider here is worth as much.

So while it listens, the Play tab says which row of that table you are standing in — good spot, workable, marginal, or too noisy. It is deliberately a word and not a number: the reading behind it is reliable for ranking one spot against another, but the same true headroom reads about 23 on a busy beach and about 30 on a thin close-mic recording, because the background sets where the noise floor settles. Printing that as decibels would be inventing precision. The bands are set by what the reading actually predicted across both sets of tests, so the word is honest even though the number would not have been.

Four recordings of people actually playing — 137 seconds, 265 hits marked by hand. This is what the settings are now fitted to, and what they achieve on it. Their headroom is written beside each one, because that is what the differences between them are.

Across all four it counts 81% of the hits, and 80% of what it counts is real. But three of the four are close-mic clips filmed for an audience, sitting at 36 to 45 dB, which is not where a phone on a towel sits. The fourth is two carbon paddles in wind at 26 dB, and that one is the realistic case. All four moved this week — 73% to 81% overall, and the windy one 58% to 72%.

Read those figures with care. The settings were fitted to these same four recordings, so they flatter themselves. Refitting on three and scoring on the fourth — four times over, each recording held out in turn — gave between 40 and 55%, and four recordings is too few to say where in that range. The two changes that produced this week's gain are different: both were fitted on built beaches and then applied to these recordings without a number being touched, so those particular improvements are not self-flattery. The absolute figures still are. On your beach, expect rather less than the table, and expect the position of the phone to be what decides it.

The first thing worth measuring was which half was losing the hits. It turned out not to be the half everyone tunes: the onset detector was finding 99 to 100% of real hits at every headroom, even at 10 dB. Every missed hit was found and then thrown away. Two things came out of that.

1. Listen in seventeen narrow bands, not five wide ones. The old onset detector summed the rise across five octave bands and fired nineteen times a second on a beach — a dozen false alarms for every real hit, which is why everything downstream had to lean so hard on loudness. Seventeen third-octave bands rising inside the same two milliseconds is a much narrower claim: a swell lifts a stretch of the spectrum, a ball lifts all of it at once. That cut the candidates by three quarters and took the four real games from 73% to 81% counted at the same precision.

2. Learn what your own paddle sounds like. The test that threw away the most real hits was the loudness one, and it is the least informative here — it describes the beach, not the ball. So it now keeps two sets of pictures: every hit it has counted, and every sound it threw out on shape. Both come from your microphone, your paddle and this minute, so a quiet sound that fails on loudness alone gets one more question — is it closer to the hits, or to the beach?

Calibrating gives the second one a head start: the bank begins with your own hits instead of waiting for six to get through on their own. That matters most exactly where waiting hurts — in a windy game the plain tests pass about one hit in eight, so a minute goes by before the bank is worth consulting. Seeded with 25 calibrated hits, 16 dB went from 19% to 28%, with fewer false counts as well. Your paddle is kept between sessions in about nine kilobytes.

What did not work, and why it is worth saying. The obvious move was a classifier trained offline on real hits replanted into built beaches. Held out by recording, it scored 51% against the hand-set tests’ 40% on more built beaches — and 17% against their 57% on the four real recordings. It had learned what a generated beach sounds like. Also measured and discarded: narrowing the bands used to measure the hit rather than to find it, refitting the limits (a gain on built beaches that reversed on real ones), letting the banks overrule a test of shape, and making the limits depend on the headroom. That is why this app contains no trained weights: held-out synthetic scores turned out to be no evidence at all about real audio, and comparing sounds within one recording is the one kind of learning with no domain to cross.

A matkot hit arrives in under three milliseconds, is twenty decibels down inside sixty, is gone ninety milliseconds later, and lifts every band from 800 Hz to 11 kHz at the same instant. Waves, wind and voices fail the decay; footsteps and doors fail the rise and sit too low; bottles ring far too long. A hit on its own stays provisional until a second one follows, which is what keeps the pair playing down the beach out of your count.

Two earlier versions of this set its limits from invented hits — once chosen by ear, once from plate theory. Both were wrong, in opposite directions, and against real play the second counted one hit in five. Tuning against recordings of people actually playing is what fixed it.

Keeping the wrong things out is still checked against forty simulated beaches carrying 5,280 noises — claps, bottles, doors, footsteps, keys, hard consonants, dogs, gulls, children and somebody's speaker — because real recordings of a good rally contain almost none of those. 86% of what it counts there is a real hit.

One of those beaches is built entirely from real recordings rather than models — applause, breaking glass, an explosion, a storm, a train, plucked strings, a crowd, dropped objects. It counts none of them, and real recordings are layered into all forty of the others too. What no test here can settle is how long your own paddle rings, which is what calibration measures, and what the note under the counter will tell you if it is the thing getting in the way.