Play a note.
Read the cents.

The instruments

Real strings, real targets

Choose an instrument and its open strings appear as a row, each with the frequency it should be at — computed from your own A4, not read out of a table, so the whole row moves when the calibration does. While the tuner is listening the nearest string lights up and shows its deviation, so tuning a guitar is six notes and no menu.

Press a string to hear it. The reference tone is an oscillator with a short attack and a short release, so it starts and stops without a click, and it holds for twenty seconds unless you press it again. If the microphone is open it will hear the tone as well as the instrument, so use headphones or stop the tone before you play.

Chromatic is the default and needs no instrument: it names the nearest of the twelve notes and measures against that, which is what you want for anything not on the list — a harmonica reed, a tuning fork, a saw. Choosing an instrument does one further thing, and the method section explains why: it narrows the range the detector searches to that instrument’s own, which is what keeps two strings ringing together from being read as the slower pitch they make between them.

The method

How the note is arrived at

Why not an FFT

The obvious tuner is a spectrum and its loudest bin. It fails twice, and both failures are on the instrument most people arrive with. The first is resolution: at 44.1 kHz a 2,048-point transform has bins 21.5 Hz apart, and a semitone at the guitar’s low E — 82.41 Hz to 87.31 Hz — is 4.9 Hz. More than four semitones fit inside one bin down there, so the answer at the bottom of the instrument is not a note, it is a neighbourhood. The second is the missing fundamental: a plucked low E is a weak first harmonic underneath a much stronger second and third, so the loudest bin is routinely the octave or the twelfth, and a tuner reading it will tell you a perfectly tuned string is an octave sharp.

The period, not the peak

Autocorrelation asks a different question: at what delay does this signal most resemble itself? That delay is the period, and the period of a note is the period of its fundamental whether or not the fundamental is the loudest thing in it. The measure used here is the normalised square difference — McLeod’s function — which divides by the energy actually present at each delay. Plain autocorrelation falls off as the delay grows for no musical reason, because fewer samples overlap, and that bias pushes every reading towards the shorter period; normalising removes it and puts the result on a fixed scale from zero to one, which is what makes a clarity threshold mean something instead of being a number somebody tuned by ear.

The window, and the refinements

The window is 8,192 samples — 186 ms at 44.1 kHz, about five and a half periods of a five-string bass’s low B at 30.87 Hz. Twenty of them are measured a second. The winning delay is then interpolated with a parabola through its two neighbours, because one whole sample of delay is 3.2 cents at the guitar’s low E and 12.9 cents at the E two octaves above it, and a tuner that cannot resolve better than thirteen cents is a tuner you cannot tune to. And the delay taken is the FIRST peak that clears ninety per cent of the tallest, not the tallest itself: the tallest peak of this function often sits at a multiple of the true period, and taking it is the other way to report an octave that was never played.

When it declines to answer

Two gates, and both of them produce a sentence rather than a reading. A frame quieter than −46 dBFS is refused outright: with no signal an autocorrelation still returns its best guess, and that guess wanders and looks exactly like a string drifting. A frame whose clarity is under 0.6 is refused as well — a room, a breath, a knock on the body of the instrument are not pitches, and the honest answer is to say nothing rather than to name the loudest thing present.

Two strings at once

This is worth stating plainly, because it is where every autocorrelation tuner is weakest and the failure looks like a bug. Two notes sounding together are not two pitches to this method — they are one signal, and that signal really is periodic, at the slower rate the pair repeats at. An E and an A together repeat about 27.5 times a second, so the mathematically correct answer to “what is the period of this window” is a note neither string is playing. Nothing in the arithmetic is wrong there; the question was.

So when an instrument is chosen the page hands the detector the range that instrument’s strings actually live in, a minor third either side. Most of those composite periods then fall below the range and are refused rather than reported, which is why the panel says it cannot hear a pitch instead of naming one nobody played. It is not a cure: where the two notes make a simple ratio the composite lands inside the range anyway — a B and the E above it are three to four, and repeat at exactly the E two octaves down, which is a guitar string. Chromatic mode has no range to narrow at all. Either way the rule is the same one every tuner has: play one note at a time, and let the last one stop ringing.

The needle

What the needle points at is the median of the last five accepted readings. A median of an odd-length list is one of the values in that list, so the number on screen is always a number some frame actually measured — an average is not, and an average is what puts a tuner between two notes when one frame reports the octave. Between two of those targets the needle eases, which is animation from one measurement to the next; it never extrapolates, never decays towards zero, and does not move at all while nothing is arriving. When the sound stops, the last reading stays on screen, dimmed, and says it is being held.

Where the audio goes

Nowhere. The stream from getUserMedia reaches an AnalyserNode and stops there; the analyser is not connected to the speakers, because connecting a microphone to the speakers is feedback. Each window is copied into one reused array, measured, and overwritten by the next one. There is no fetch, no XMLHttpRequest, no WebSocket and no upload target in any of this property’s own code — the three scripts it adds are js/pitch.js, js/mic.js and js/tone.js, with js/app.js for the page, and you are welcome to read all four.

One thing on this page does reach the network, and it is worth naming rather than leaving for you to find in the developer tools: the site’s Google Analytics tag, which every page on labs.llc carries. It loads in the head, records that a page was viewed and sets its own cookie. It is nothing to do with the tuner, it is not passed a single sample, and it runs whether or not you ever press Listen.

If the microphone will not open

Every way that can go is answered in words on the panel above rather than by a dead needle: a page not served over HTTPS, where browsers do not offer the microphone at all; a browser without the API; a refusal, which will not prompt again until you change it at the padlock; a prompt dismissed without an answer, which will; no microphone on the machine; one held by another application; and one unplugged in the middle of a session. The difference between a refusal and a dismissal is read from the Permissions API where the browser offers it, and where it does not, the page says both possibilities instead of picking one.