The microphone is not used until you press that. When you do, the audio is analysed frame by frame in this page to find the pitch, and then discarded. It is never recorded, never stored, and never sent anywhere — there is no upload in this tool and nothing to upload to. Stopping releases the microphone immediately, and your browser lets you revoke the permission at any time.
| String | Note | Frequency | Play |
|---|
Everything above was verified against steady synthesised tones, and it passed: a fifth of a cent, no octave errors. Every real complaint about this tool was about something that test cannot see — what happens across the life of a note, and how the number behaves while a peg is turning. So a second harness synthesises a plucked string: exponential decay, partials slightly sharp of true harmonics because a real string is stiff, upper partials dying faster than the fundamental, room noise underneath. It runs the page's own code frame by frame at display rate. Every complaint turned into a number on the first run.
| what it does | before | now |
|---|---|---|
| a plucked low E stays readable for | 1.9 s of an 8 s note | the whole 8 s |
| …covering | 12% of the note | 100% |
| display blanks and returns, six strings | 57 times | never |
| overshoot after a 40-cent peg turn | 5.8 cents | none |
| …and it then settles | 5.0 cents wrong | 0.2 cents wrong |
| displayed error on a steady low E | 3.8 cents | 0.1 cents |
| CPU per frame | 7–12 ms of a 16.7 ms budget | 1.1 ms |
| a guitar played 20 dB quieter | 16% of a low E readable | 100% |
The four faults were not one fault. The correlation ran at full sample rate over every lag down to 58 Hz — three and a half million multiply-adds a frame, most of a 60 fps budget, so the browser dropped frames, and dropped frames are lag. The queue of recent readings lost an entry for every rejected frame, so three misses in a row emptied it and the display blanked. The clarity needed to keep a note was fixed, while the clarity of a real note falls steadily as its harmonics die. And the smoother was a median in front of a nonlinear average, which is a lag followed by something that keeps moving after the input stops.
Each got its own fix, and one of them is worth stating on its own: a peak that lands where the pitch already was is evidence in itself. Once a note is being tracked, a peak within a semitone of it is accepted at a sixth of the clarity a cold start needs — and the octave error that a low threshold usually buys is precisely what that agreement test refuses. That is what lets a note stay readable long after its harmonics have gone, which is the part of the note you spend turning the peg.
Measures how long the waveform takes to repeat. McLeod: correlate the signal with a delayed copy of itself, take the shortest delay nearly as good as the best, interpolate between samples.
On a plucked low E the loudest partial is usually not the fundamental, and small microphones roll off below 100 Hz. Correlation asks when the whole waveform repeats, which does not need the fundamental to be present.
Forwards and backwards, so no phase shift. Mains hum, air conditioning and handling noise sit below the lowest guitar note and cost up to 25 cents on a low E.
A coarse pass finds the pitch, the signal is band-passed around it to isolate the fundamental, and the peak is measured there. The stages must agree or the tool says nothing.
Frets are placed for twelve equal semitones, so every key sounds the same and none is exactly right. A major third in equal temperament is 14 cents sharp of the pure interval your ear wants.
Fretting stretches the string, so a guitar tuned perfectly open plays slightly sharp everywhere else. Individually adjustable saddles spread that error rather than remove it.
Machine heads and nut slots have slack in them. Arriving from below leaves the string under tension against the post, so it holds; arriving from above leaves it to creep flat as you play.
Two notes a few cents apart make the volume pulse, once per second per hertz of difference. A tuner is faster, but beats tell you the interval is right rather than each note separately.
Nylon and freshly wound steel keep stretching for hours. Tune, play, tune again, several times, before deciding the instrument has a problem.
With a second string ringing through a held note the reading used to move 461 cents and take over a second to recover; a clap moved it 0.3 and a voice 3.5, so the problem was other musical sounds, not noise. The tracker now carries a window that travels with the pitch, up to 600 cents a second (about four times a brisk peg turn); anything outside is not the note. A new pitch is adopted only after five consecutive frames, clean (clarity about 0.98) and agreeing within 25 cents, whereas two notes together read 0.55–0.64 and wander tens of cents. A real string change is picked up in a tenth of a second.
A one-euro filter: first-order, so it is a weighted average of the past and cannot overshoot; its cutoff rises with the speed of the reading, so it sits still on a held note and follows a turning peg. It replaced a median feeding a variable-weight average: on a 40-cent peg turn that sailed 5.8 cents past and settled 5.0 cents wrong. Now there is no overshoot, reaching the new pitch two hundredths of a second after the peg stops.
The old single gate rose to meet the note and shut behind it. Now you must be clearly above the room to start a reading (about −64 dB relative to a firm pluck) and barely above it to keep one (down to −84 dB), a gap of 20 dB; the floor learns only from frames where nothing was found. On a two-second decay that is five more seconds of usable reading, the part you spend turning the peg.
On steady synthesised tones, on the reading the tool actually displays: 0.03 cents median on a clean signal and 0.24 with mains hum, rumble or handling noise; worst case 0.3 clean and 2.5 contaminated, across 75 cases. Most people cannot hear 5 cents. It refuses to read rather than guess, judging frames by their scatter around a fitted line, so a turning peg reads fine and jitter refuses: 5 of 30 contaminated cases decline.
Almost always a second string is still ringing. With an adjacent open string sounding the tool declines in 23 of 25 cases and the two it reads land about a third of a semitone out. That is a real limit, not a bug: damp the other strings and it resolves at once. Naming a string narrows the search to a minor third either side.
Searching narrowly around the tracked note before looking at the room changed no number in the test; tightening the band-pass kept a neighbouring string out but made speech four times worse. Measuring them showed they earned nothing, so they went.
Measured against synthesised tones of known frequency, on the reading the tool actually displays rather than on a single frame — and, separately, against a synthesised plucked note with things going on around it: another string ringing through it moves the reading 0.2 cents, a clap 0.3 and a nearby voice 3.5, all recovering immediately. On steady tones: 0.03 cents median error on a clean signal and 0.24 cents with mains hum, air-conditioning rumble or handling noise added, worst case 0.3 cents clean and 2.5 cents contaminated, across 75 cases. For comparison, most people cannot hear a 5-cent error, and a fretted guitar is further out than that by design.
It refuses to read rather than guess — but it refuses on the right grounds. The test used to be the total spread of the recent frames, which meant that turning a peg looked exactly like a fault: a steady climb of forty cents is a large spread. The frames are now fitted with a straight line and judged on how far they scatter around it, so a peg being turned reads fine and jitter still refuses. A very noisy signal still sometimes says nothing at all — 5 of those 30 contaminated cases decline.
Because the gate that decides whether there is a note was a single threshold, and a plucked string spends most of its life below it. Three things were wrong at once: the quietest signal accepted was fixed, a second clarity test in the display code was stricter than the one in the detector, and — worst — the adaptive floor learned from the note itself, so playing loudly raised the very threshold the decay then had to clear. The gate rose to meet the note and shut behind it.
There are now two thresholds. You have to be clearly above the room to start a reading and barely above it to keep one, and the floor only learns from frames where nothing was found. Measured on the shipped detector: acquiring needs about −64 dB relative to a firm pluck and holding runs down to −84 dB, a gap of 20 dB. On a string decaying with a two-second time constant that is five more seconds of usable reading — which is the part of the note you spend turning the peg.
Almost always because a second string is still ringing. Two notes together have their own repeating period, and it distorts the answer even when the search is narrowed to one string: on tests with an adjacent open string sounding, the tool declines to read in 23 of 25 cases and the two it does read land about a third of a semitone out. That is a real limit, not a bug to work around — damp the other strings and it resolves immediately.
No. The microphone is only opened when you press start, the audio is analysed in this page and discarded frame by frame, and nothing is recorded, stored or transmitted. There is no upload in this tool. Stopping releases the microphone, and the permission is your browser's to revoke whenever you like.
It does not have to — leave it on auto and it will name whatever it hears. But if another string is still ringing, the two notes together have their own repeating period, and a pitch detector that reports it is not making a mistake: the sum really does repeat at that rate. Naming a string narrows the search to a minor third either side, which is the only way to answer "how far is this string out" while another one sounds.
Intonation. Pressing a string down stretches it slightly sharp, and the amount depends on the action, the gauge and where the saddle sits. That is why saddles are set individually, and why a guitar tuned perfectly open is never perfectly in tune everywhere at once.
440 Hz unless something else in the room says otherwise. Orchestras often sit at 442 or 443, some recordings are at 432, and baroque ensembles commonly use 415 — a semitone flat, which is why the tool goes down that far.
At A = 440 Hz the six open strings are E2 82.41, A2 110.00, D3 146.83, G3 196.00, B3 246.94 and E4 329.63 Hz. Each is 440 × 2s/12 Hz, where s is the number of semitones above or below A4; the low E is 29 semitones down.
Releases in which this page changed, newest last. Derived from the archived copy of every release, not from notes written afterwards — so it reflects what actually shipped. Site-wide passes are left out; they are in the full changelog.