Enough to put every sustained note within about 10 cents of its target, and no more. A trained listener can hear a difference of 5 to 6 cents; most adults reliably notice 25 (Wikipedia)). Inside that window a note reads as sung. Outside it a note reads as wrong. On a lead I correct pitch centres to 80 percent in Melodyne, leave the vibrato alone, reduce slow drift by no more than a third, and keep a retune speed of 0 ms for records where the tuning is meant to be heard.
How many cents off can a note be before somebody hears it?
Five or six for a trained ear on isolated notes, and the threshold moves with frequency, level and timbre (Wikipedia)). Two other numbers matter more in a mix. In one study, a change of tone quality made student musicians worse at recognising notes that were 12 cents out; and adults without amusia recognise 25-cent differences very reliably. Tonal context sharpens judgement, per the same source, which is why a note that passed in solo sounds flat once the pad comes in.
The working limits in my room: a lead over a moving arrangement sits within 10 cents of centre on sustained notes; over a held chord, within 5; a note shorter than 100 ms gets no correction, because nobody resolves its pitch anyway. The exception is a note that lands more than 25 cents out and stays there. That one gets moved whatever its length, because it reads as a wrong note rather than a flat one.
What is being corrected: centre, drift or vibrato?
Three things, and Melodyne separates them. The pitch centre is the average pitch of a note; the Correct Pitch macro moves centres toward the nearest semitone, or toward the current scale or chord if Snap to Chord Scale is on, and at lower settings it moves only the notes that are far out while leaving the near ones alone (Celemony). Pitch drift is the slow wander across a held note, which Celemony describes as symptomatic of poor technique, and the macro's second slider reduces it without touching faster movement. Pitch modulation is the fast, intended movement, vibrato and trills, and it has its own tool, with 100 percent meaning the original and 0 a flat line (Celemony).
That split is the whole method. A singer's vibrato is measured in tens of cents: a study of professional performances of Schubert's Ave Maria found spans between plus or minus 34 and plus or minus 123 cents, mean plus or minus 71, and the evidence is that listeners hear the mean frequency as the pitch (Wikipedia)). Correct the centre and the vibrato rides along untouched. Flatten the modulation and the same note sounds like a synthesiser, however accurate its centre.
What are my settings on a lead, and in what order?
Comp first, tune second: the note separations are checked, then the macro runs; my notes on comping a lead in two passes cover the first half.
The macro pass: pitch centre 80 percent, pitch drift 30 percent, Snap to Chord Scale on once the chords are entered. At 80 a note that was 40 cents out ends up 8 out, inside the window; a note that was 5 out barely moves. Then the manual pass, note by note. Scoops into a note longer than a beat get the drift tool pulled to 50 percent, so the note arrives sooner. Phrase endings that fall 30 to 50 cents in the last 100 ms stay as sung, because that fall is speech and the ear expects it. Fast ornaments, the three-note turns that Ukrainian folk-style singing puts at the top of a phrase, get checked before the macro runs: the detection sometimes reads a 200 ms turn as one note with modulation, and the macro then straightens the ornament into a single pitch. Separating it into three notes by hand first keeps the turn.
Auto-Tune does a different job here. Its Retune Speed is the time in milliseconds the pitch takes to reach the target; Antares recommends 20 to 50 ms for most vocals, 40 to 80 for slow lines and 10 to 20 for fast passages, and calls 0 to 10 ms the effect (Antares). It goes on a demo for real-time correction, or on a lead when the effect is the point; a finished lead gets the offline pass, which can leave the ornament alone and move the long note next to it.
How are doubles and harmonies tuned differently?
Doubles get less correction than the lead; harmonies get more, and to a different reference. Two voices 10 cents apart on A3 beat at about 1.3 Hz, since 10 cents at 220 Hz is 1.27 Hz, and that beat is the width of a stacked pop chorus. So doubles are tuned to the lead, with the centre at 60 to 70 percent, so a few cents of spread survive; take them to 100 and the double folds into the lead and the stack narrows.
A sustained third that is 10 cents off its chord beats against the pad and the piano; in harmony, as the same source puts it, very small changes cause large changes in beats. Harmony centres therefore go to 100 percent, drift to 50, and Snap to Chord Scale on. One refinement on held chords: an equal-tempered major third is 400 cents and the just third is 386, so a stack tuned exactly to the grid rings like a piano. Dragging the third of a held chord 8 to 10 cents under the grid in Melodyne makes a three-voice stack lock like a choir. Only on the last chord of a chorus, because the piano underneath stays equal-tempered.
When is the tuning meant to be heard, and how is it set?
When the record says so. Cher's "Believe" in 1998 is the usual reference, made by setting Auto-Tune's retune speed close to zero so that corrections snap instead of glide (Antares). Three settings decide whether the effect works. Retune Speed at 0 to 10 ms. The scale set to the song's actual notes, because at 0 ms a wrong scale is a wrong note, loudly. And formant correction on for any move over two semitones, so that a voice pushed up a third keeps its size.
The effect gets printed to a new track and the untuned lead stays in the session: the artist who wants the effect on Monday wants it halfway off on Thursday, and the printed track under the dry one at minus 6 dB is halfway.
What changes when the song will be sung live on television?
The correction on the record gets tighter than the stage will manage, but only slightly. A contest performance is sung live, and if the studio lead has been pulled to 100 percent everywhere, the broadcast sounds like a different singer. Keeping the record at 80 percent centre, with vibrato and phrase endings untouched, leaves a version a good live take can match. The note on a contest version covers which versions exist. The records I have mixed and mastered are listed with sources at /credits.
Frequently asked
How many cents off can a vocal be and still sound in tune? Within about 10 cents of the target on a sustained note over a moving arrangement, within 5 over a held chord. A trained ear hears 5 to 6 cents on isolated notes; most adults reliably notice 25, where flat turns into wrong.
Melodyne or Auto-Tune for natural vocal tuning? Melodyne for a finished lead, because it separates pitch centre, drift and vibrato and lets a fast ornament stay untouched next to a corrected long note. Auto-Tune for real-time correction on a demo, or when the audible effect is wanted, at a Retune Speed of 0 to 10 ms.
Can a listener tell when a vocal has been tuned? Rarely from centre correction at 80 percent, because the vibrato and the phrase endings stay human. Flattened modulation, an ornament averaged into one pitch, or a snapping retune speed give it away.
If a lead needs tuning without sounding tuned
Send the comped lead as a dry WAV, the key and a chord chart with bar numbers. The tuned lead comes back as a separate file, with a list of every note moved more than 15 cents and its bar. Background and contact at /about.