Why tuned vocals sound robotic, and how to avoid it
Everyone can hear when a vocal has been tuned badly. Fewer people can say what the actual defect is — and the answer tells you how to avoid it.
Pitch and formants are different things
Your vocal cords produce a pitch. Your throat, mouth and nose then act as a resonant filter on top of it, emphasising certain frequency regions. Those resonances are called formants, and they are what make a voice sound like a particular person, and what makes "ee" sound different from "oo".
Critically, formants stay roughly where they are when you sing higher or lower. Your mouth does not change size when you go up a fifth. So pitch moves and formants mostly do not.
What naive correction gets wrong
The simplest way to change pitch is to resample — play the audio faster or slower. That moves the pitch, and it drags the formants with it by the same amount.
Move a voice up a few semitones that way and the formants go up too, which your ear reads as a physically smaller person. That is the chipmunk effect. Move it down and you get the opposite: an unnaturally large, hollow sound.
Even at small correction amounts this is audible as a subtle wrongness — the take is in tune and no longer sounds like the singer.
This is why heavy correction on a take that was already close often sounds worse than light correction on a take that was further off. The artefact scales with how much you moved things, not with how wrong the original was.
Formant-preserving correction
The fix is to separate the two: shift the pitch while holding the formant structure where it was. Phase-vocoder and time-domain algorithms that do this properly — rubberband is a well-known one — keep the voice recognisable through much larger corrections.
The second thing that matters is how fast the correction happens. Snapping instantly to the target pitch removes the natural scoops and vibrato that make singing sound like singing. That instant snap is the deliberate Auto-Tune effect; if you do not want the effect, you want a slower retune speed that lets the performance breathe.
- Preserve formants — this is the single biggest factor
- Do not snap instantly unless you want the effect
- Correct the notes that are wrong, not the whole take
- Fix the performance first; correction is not a substitute for another take
Doing it in dhun
Pitch is detected with CREPE and corrected with rubberband, so formants are preserved by default. You do not have to know to ask for that.
You can also ask it to list what you actually sang — note by note, with how far off each one was — and then fix only the ones you name. "The third line is flat" is a complete instruction; it will find the line, tell you which note and by how many cents, and correct just that.