How pitch detection works
A sung or played note is a waveform that repeats. This detector uses autocorrelation: it takes a short window of audio (2048 samples, about 43 ms at 48 kHz), slides a copy of it against itself, and measures how well the two line up at every offset. The offset where they match best is the length of one cycle, the period. Interpolating between neighbouring offsets gives a period finer than one sample, and the frequency is simply the sample rate divided by that period: a 100-sample period at 44,100 Hz is 441 Hz. Windows that are too quiet, or that never line up cleanly, are reported as no pitch.
In the Audio file tab the window steps through the file every 1024 samples (50% overlap), so a three-minute song becomes several thousand readings. The work is done in small slices between screen updates, which keeps the page responsive on a long file; you can stop part-way and keep what has been analysed.
Pitch vs frequency
Frequency is physics: how many times per second a waveform repeats, measured in hertz. Pitch is perception: how high or low the sound seems to a listener. For a steady musical tone the two track each other closely, which is why a detector can name a note from a frequency, but they are not identical. Your ear hears the pitch of a cello note even when the fundamental is weak, by inferring it from the spacing of the harmonics, and very loud or very short sounds can seem slightly higher or lower than their frequency suggests.
Notes are named by fixing one reference, normally A4 = 440 Hz, and dividing each octave (a doubling of frequency) into twelve equal steps. Change the A4 box above to 415 Hz for baroque pitch, 442 Hz for many European orchestras, or 432 Hz to compare, and every note name and cents reading shifts with it. The note to frequency converter does the same maths for any single value.
Cents, explained
A cent is one hundredth of a semitone, so there are 1,200 cents in an octave. Because the scale is logarithmic, a cent is a different number of hertz at every height: about 0.25 Hz near A4 but only 0.06 Hz near A2. That is why tuners and this detector report cents rather than hertz for "how far off". The readout always measures from the nearest note, so it runs from −50 (flat) to +50 (sharp); past 50 cents the nearest note changes.
The needle turns green within ±5 cents, the usual tolerance for an in-tune note. Healthy vibrato swings a few tens of cents either side several times a second, so a sung note that never sits still on the needle can be perfectly musical; the pitch graph shows whether the wobble is centred on the note.
What people use it for
Singing practice. Sing a phrase and read the graph: scoops into notes, drift on long notes and a flat top note are all visible. For drills against a target note, the vocal pitch monitor adds an in-tune band, and the vocal range test finds your lowest and highest notes.
Transcription. Hum or whistle a melody you cannot place, or drop in a recording of a solo line, then click along the graph to hear each note in isolation and read its name. The most-common-notes list is a quick hint at the key.
Tuning samples and sound design. Find the root of a one-shot, a vocal chop or a synth patch before you map it to a keyboard, and see how many cents to retune it by. Export the CSV to check a whole sample pack in a spreadsheet. To shift a file once you know the offset, use the pitch shifter.
Tuning an instrument. The live readout works as a tuner, but the chromatic tuner is built for it, with a longer listening window for low strings.
Honest limits
- One note at a time. Chords, full mixes and two singers at once confuse autocorrelation; it reports whichever pitch dominates or nothing.
- Reverb and noise. A long echo lets the last note ring under the next one, and steady background noise can mask quiet notes. A close microphone and a dry room help most.
- Range: 50 Hz to 2,000 Hz. Very low bass notes and the top of the piccolo or piano are outside it, as are unpitched sounds like drums, hiss and most consonants.
- Octave slips. Occasionally a reading jumps exactly one octave; the note name stays right.
- Long files. The first 10 minutes of a file are analysed. Decoding very large files can run out of browser memory on phones; cut out the part you need with the audio trimmer first.