Pitch detector

How to read the pitch detector

Press start, allow the microphone, then sing or play one note and hold it. The large letter is the nearest note in equal temperament, with its octave number: A4 is the A above middle C, and middle C itself is C4. Under it you get the frequency the detector measured, and the target frequency of the note it is closest to. The needle and the number beside it show the gap between the two in cents. When the needle sits in the green band you are within five cents of the note either way — a deliberately tight margin, since tuning is about getting the needle to the middle and keeping it there.

Below the readout, the trace draws the last eight seconds of pitch against a grid of note lines. It is the more useful half of the tool for singers, because a held note is never a single number: it starts a little under, settles, wobbles with vibrato and sags as the breath runs out. The needle shows you this instant; the trace shows you the shape.

Cents, hertz and the note name

Frequency is the raw measurement: how many times per second the sound repeats. Pitch works on ratios rather than differences, so every octave is a doubling — A3 is 220 Hz, A4 is 440 Hz, A5 is 880 Hz — and the twelve semitones inside an octave are each a step of the twelfth root of two, about 5.9 percent. That is why a fixed error in hertz means very different things at different heights: 2 Hz off at 82 Hz (a guitar’s low E) is about 42 cents, a clearly sour note, while 2 Hz off at 880 Hz is under 4 cents.

Cents fix that by measuring on the same ratio scale the ear uses. There are 100 cents in an equal-tempered semitone and 1,200 in an octave, so “+12 cents” means the same amount of sharpness whether you are playing a bass or a piccolo. A reading can never be more than 50 cents from the nearest note: past that point it is closer to the next note, and the letter changes instead.

How pitch detection actually works

The obvious approach — take a spectrum and report the loudest frequency — fails on almost every real sound. A voice or an instrument produces a fundamental plus a stack of harmonics at two, three, four times that frequency, and the fundamental is often not the loudest of them. On a sung vowel, whichever harmonic happens to fall near one of the vowel’s resonances can be much louder than the fundamental, and a spectrum-peak detector would call the note an octave or more too high.

What stays constant is the period: however the harmonics are balanced, the combined waveform repeats once per cycle of the fundamental. So this detector looks for repetition in time instead. It takes the most recent 2,048 samples — about 43 milliseconds at 48 kHz — and compares the wave with shifted copies of itself. The shift at which the copy lines up best is the period, and the sample rate divided by the period is the frequency. That family of methods is autocorrelation; the variant used here is the McLeod pitch method, published by Philip McLeod and Geoff Wyvill in 2005, which normalises each comparison so it scores between minus one and one regardless of how loud the note is. Its close relative YIN, from Alain de Cheveigné and Hideki Kawahara in 2002, is the other standard reference in the field.

Two details make the number precise. The best alignment almost never falls exactly on a whole sample, so the detector fits a parabola through the best match and its two neighbours and reads the peak between them; on a clean test tone that puts the reading within a fraction of a cent. And the height of that peak becomes the clarity figure under the readout: a steady, periodic sound scores close to 100 percent, while hiss, breath and room noise barely line up with themselves at any shift and score low.

Why the display sometimes goes blank

A reading is shown only when clarity is above 90 percent for several frames in a row. Below that, the detector still produces a number — it always finds some best shift — but the number is a guess, and a tuner that flickers between guesses is worse than one that says nothing. The note dims for a moment when clarity drops, then disappears. The usual causes are breathy or whispered tone, the attack of a plucked string (which is noisy for the first instant), the tail of a note as it dies away, background noise that is louder than you think, and more than one note at once.

That last one is a hard limit, not a tuning problem. This is a monophonic detector: it is looking for one repeating period. A strummed chord or two people singing contains several periods at once, and the answer it settles on, if any, will not be meaningful. Tune one string at a time.

Octave errors, and why an instrument can read an octave out

A waveform that repeats every period also repeats every two periods, and every three. Choosing between them is where pitch detectors go wrong, and the classic failure is an octave error: reporting a note twelve semitones away with the right letter name. The McLeod method guards against the low-side version by taking the first strong repetition rather than the strongest, and the detector refuses to guess when the period is shorter than its 2,000 Hz ceiling rather than quietly reporting the octave below. Very low notes with a weak fundamental are the remaining risk.

The detector listens from 50 Hz up. That covers every singing voice — the textbook bass range starts at E2, about 82 Hz — and the fundamentals of most instruments. The lowest string of a four-string bass guitar, E1 at about 41 Hz, is below it. Tune that string using its 12th-fret harmonic instead, which sounds E2, an octave up and well inside the range.

Tuning an instrument with it

Pluck or bow firmly, then watch the needle once the attack has passed and the note is ringing steadily; the first instant of a plucked note is noisy and can read slightly sharp. Bring the string up to pitch from below rather than down from above, so the gear or peg is holding tension rather than resting against slack. For reference, at A = 440 Hz the target frequencies are:

• Guitar, standard tuning: E2 82.41 Hz, A2 110.00, D3 146.83, G3 196.00, B3 246.94, E4 329.63.
• Bass guitar: E1 41.20 Hz (use the harmonic, as above), A1 55.00, D2 73.42, G2 98.00.
• Violin: G3 196.00 Hz, D4 293.66, A4 440.00, E5 659.26.
• Ukulele, standard re-entrant tuning: G4 392.00 Hz, C4 261.63, E4 329.63, A4 440.00.

A detector like this measures the fundamental, which is what a tuner should do, but it does not know about stretch tuning. Pianos are deliberately tuned with their top octaves a little sharp and their bottom octaves a little flat of the equal-tempered figures, because stiff piano strings produce harmonics slightly above the exact multiples. A piano that reads a few cents off at its extremes may well be correct.

Choosing a reference pitch: 440, 442 and 415

Every note name on this page is measured against the A you set. A = 440 Hz is the international standard: adopted by the International Organization for Standardization as a recommendation in 1955 and formalised as ISO 16 in 1975, and the default on electronic keyboards and tuners. Many orchestras tune a little higher; in continental Europe A is commonly anywhere from 440 to 444 Hz. And period-instrument ensembles have settled on 415 Hz as a modern “Baroque pitch”, very nearly a semitone below 440 — which is why the control stops at 415 at the bottom and at 466 at the top, the A-sharp above.

If you are singing or playing along with a fixed-pitch instrument, match that instrument, not the standard. The quickest way is to play its A into this page at 440, read the frequency, and type that in as the reference. Your setting is remembered in this browser for next time.

Intonation for singers: what the needle can and cannot tell you

The needle measures against equal temperament, the tuning of a piano and a guitar, where every semitone is exactly 100 cents. Unaccompanied singers and string players often do not tune that way, and are right not to. A pure major third — the frequency ratio 5:4, the interval that sounds still and ringing in a sustained chord — is about 386 cents, nearly 14 cents narrower than the equal-tempered 400. A pure fifth, 3:2, is about 2 cents wider. So when a choir’s thirds read “flat” on a tuner, that may be good ensemble tuning rather than an error. Use the needle to find your reference and check single held notes, not to police every interval in a chord.

Vibrato is the other trap. A healthy vibrato is a regular swing of pitch either side of the note several times a second, and a needle following it will never sit still. Read the trace instead: the note you are singing is the centre line of the wave, not its peaks. A trace that drifts steadily downward over a long note is the pattern to work on — usually breath support running out before the phrase does. To see which notes your voice can reach at all, the vocal range test uses the same detector to find your lowest and highest held notes and place them on a keyboard.

Getting a clean reading

• One sound source. Turn off the backing track while you tune, or wear headphones: the detector cannot tell you from the speakers.
• Close to the microphone. A hand-span away is plenty for a voice. A laptop microphone across the room hears as much of the room as of you, and clarity falls.
• Processing off. The page asks the browser to switch off echo cancellation, noise suppression and automatic gain, which exist for calls and can gate or bend a held note. Some systems apply their own processing anyway — if readings behave oddly, the microphone test shows which of those settings your browser says it applied.
• A quiet room. Fans and air conditioning are broadband noise that lowers clarity for every note. The decibel meter will show you how much there is.

Nothing is played back through your speakers, so there is no feedback to worry about. If you need a steady reference tone to sing against, the tone generator in the sound test will hold any frequency you type, which pairs well with this page on a second device.

Where VoiceSnap Pro fits

This detector is a side project of a company that builds dictation software, where the microphone is the whole product. VoiceSnap Pro is a voice-to-text app for macOS and Windows: hold one keyboard shortcut, speak, and clean punctuated text appears in whatever field your cursor is already in, and every dictation is saved to a searchable notes library. It is a one-time $39 purchase, not a subscription.

It has not shipped yet, so there is nothing to download today — join the waitlist for one email on release day. Until then, the browser-based voice typing tool shows what speech-to-text makes of your voice.