Two-Pass AI Tuning That Keeps Emotion: My 2026 Vocal Workflow in Logic Pro

Horia Stan is music producer and sound engineer at The One Records, Bucharest.

The problem I stopped accepting in 2026

AI pitch correction sounds perfect. It also kills life. Clips feel glued. Breaths disappear. Micro-intonations that sell a lyric go flat. I refuse to choose between pitch and emotion. I built a workflow that gives me both. It uses Nectar 4 Assist and Melodyne together. It uses targeted manual blends. I work in Logic Pro on an Audient iD14 MkII. My plug-in palette is iZotope Nectar 4, Celemony Melodyne 5, Antares Auto-Tune X, FabFilter Pro-Q3, Waves RDeEsser, and Waves Trans-X for transients.

This is not a beginner overview. This is a specific, repeatable two-pass method that preserves +/- 10-25 cents of micro-variation while removing obvious pitch and timing problems. I list exact numbers and where to automate. You can finish a pop vocal faster and keep the soul inside.

The two-pass philosophy

I separate correction into two goals.

• Pass one: Remove gross pitch and timing errors fast with AI. Make the performance singable and consistent. Do not smooth everything.
• Pass two: Sculpt emotion manually at the note level. Restore micro-variation and breath. Choose where tuning is audible and where it is invisible.

Most people stop at pass one. That is why so many modern records sound clinical.

Why two passes work

AI is fast. It finds notes and suggests fixes. But it applies uniform correction that treats musical inflection as a mistake. Melodyne is surgical but slow if used on every note. The two-pass approach uses AI to do the heavy lifting and Melodyne plus manual blend to make the result human again.

Tools and exact settings I use

• DAW: Logic Pro 2026.
• Audio interface: Audient iD14 MkII.
• AI assist: iZotope Nectar 4 Vocal Assistant - "Vocal Assistant: Style - Modern" with Correction Strength set to 30% and Vibrato Preservation on.
• Surgical editor: Celemony Melodyne 5 - use only for final per-note moves. Keep "Pitch Drift" tolerance to retain 10-25 cents of drift.
• Hard-correct option: Antares Auto-Tune X - retune speed 15 ms for natural, 3 ms for intentional effect.
• De-esser: Waves RDeEsser - frequency 6.5 kHz, half-power Q, threshold set so sibilance peaks are -6 dB below vocal peaks.
• EQ: FabFilter Pro-Q3 in dynamic mode for the 3-8 kHz presence band - multiband dynamics set to react at 3-6 dB reduction with attack 3 ms, release 120 ms.
• Transients: Waves Trans-X to soften plosive consonants with Attack Depth -6 dB and Release 80 ms.

I name every track with the suffix -RAW and -TUNED. I keep both in the session routed to the same bus.

How I decide what to keep raw

I keep raw on:

• small portamento slides under 25 cents
• breaths and pre-phrase noise
• violent vibrato peaks that carry emotion

I tune hard on:

• notes off by 40 cents or more
• timing flubs that confuse the groove
• obvious double-tracking alignment problems

Concrete rule: if a note needs more than 40 cents correction, the TUNED track goes up for that phrase. Otherwise the RAW stays dominant.

Exact Melodyne moves I make

When I open Melodyne for a phrase I do three things only.

1. Nudge note center by up to 25 cents. I never go beyond 40 cents without checking the original.
2. Reduce Pitch Center correction for sustained notes to 50% of what Melodyne suggests. That preserves the singer's modulation.
3. Use the Formant tool only to correct obvious vowel mismatches. Typical formant moves are -3% to +3% max. More and the vocal sounds synthetic.

This yields a voice that is in tune but still breathes.

Automation details that save time

• I create a VCA-like fader for the TUNED channel and automate it. I never try to micro-fade each note in the tuned audio. I only automate phrase boundaries and problem spots.
• For chorus lift I add a static +3 to +6 dB on the TUNED track and then back it off in verses.
• I use Logic regions and markers to store "TUNE UP" regions. I work through the list and only open Melodyne on those. That saves processing power and focus time.

De-essing and transient control without killing clarity

De-essing should not be a blanket carpet. I set Waves RDeEsser to react with a threshold so sibilant peaks sit about -6 dB below the loudest vocal peaks. That keeps consonants crisp.

I use FabFilter Pro-Q3 dynamic band on 3.2 to 8.0 kHz with a 3 ms attack and 120 ms release. Reduction never exceeds 6 dB unless the vocal is harsh.

For plosives I use Waves Trans-X with Attack Depth -6 dB. That tames plosives but preserves consonant snap.

Session CPU strategy

Melodyne 5 is heavy. I only open Melodyne on frozen or bounced regions I will edit. I use Logic Track Stacks to collapse takes. I freeze VSTs I am not editing. This keeps my template under control on an Audient iD14 MkII laptop rig.

When to use Auto-Tune X instead of Melodyne

Use Auto-Tune X for speed when you need uniform tuning across a stack - for doubles or harmonies. Set retune speed 15 ms for natural, 3 ms if you want the effect. Use it on harmony submixes only. For leads I prefer the Melodyne manual pass.

Common mistakes and how I avoid them

• Mistake: tuning everything to zero cents. Fix: preserve 10-25 cents of modulation on sustained notes.
• Mistake: using formant shifts larger than 5%. Fix: keep formant moves inside -3% to +3% and check in mono.
• Mistake: over-de-essing. Fix: set target sibilance peaks -6 dB below vocal peaks and automate if needed.

Final polish and mixing notes

After tuning I treat the vocal like an instrument. I use FabFilter Pro-MB for light multiband glue only if the vocal needs it. I never compress the tuned bus harder than the raw bus. Typically I route both RAW and TUNED into the same bus and apply the same bus processing so they sit as one voice.

My vocal chain on the bus often reads: FabFilter Pro-Q3 (broad cuts), FabFilter Pro-MB (if needed), Waves CLA-2A emulation for colour, short plate reverb send, tempo-synced slap delay in parallel.

Pullquote

Concrete takeaway

Do this next session: duplicate the lead to create RAW and TUNED tracks. Run Nectar 4 Assist on the TUNED track at 30% correction. Use Auto-Tune X at 15 ms for remaining gross errors. Mark phrases that sound flattened. Open Melodyne only for those phrases and restore 10-25 cents of pitch modulation and max 3% formant moves. Automate the TUNED track fader so tuned audio is typically 0 to 35 percent of the final vocal, spiking to 50 percent only on problem notes.

That sequence gets you radio pitch with live emotion. Try it and tally how many phrases need Melodyne edits. You will cut editing time by half and end with vocals that feel human.