Dialogue cleanup guide
Strip the noise, keep the voice sounding human
AI noise removal separates speech from hiss, hum, traffic, and room tone faster than manual EQ and gating ever could. Pushed too far, though, it leaves dialogue metallic, watery, or oddly gated. This guide shows how to clean in stages and finish with a mix that sounds recorded well rather than processed hard.
From the Mellius product team · Updated September 28, 2026
Conspiracy theorists, you may hate this
AI Director
›
Just type what you need
- ›Remove all silent parts
- ›Add bold kinetic captions
- ›Only have Bill Ackman speaking
- ›Cut to highlights only
- ›Add cinematic transitions
- ›Match cuts to the beat
The short answer
Duplicate the dialogue track, apply noise reduction in light passes, and compare against the original every few decibels of reduction. Restore presence with gentle EQ, add consistent room tone under edits, then duck music beneath the cleaned voice. Stop when the scene sounds quiet, not when the meter reads silent.
Staged reduction
Light sequential passes beat one aggressive cleanup that damages consonants and breath.
A/B against raw
Every cleanup decision is checked against the unprocessed dialogue for naturalness.
Finished mix context
Cleaned voice is judged inside the full mix with music, tone, and captions, not solo.
Clean in passes, not all at once
Start by identifying the noise: steady hiss and hum respond well to AI suppression, while wind buffets, clothing rustle, and overlapping voices need editing or replacement rather than filtering. Apply reduction in two or three light passes of a few decibels each, listening to sibilants, breaths, and room decay after every pass. If ess sounds turn lispy or pauses pump audibly, you have gone one pass too far — back off and accept a whisper of background rather than a damaged voice.
Edit before you over-process. Cut coughs, mic bumps, and dead air first, because noise tools work cleaner on deliberate speech than on garbage audio. For location interviews, capture thirty seconds of room tone and lay it under the whole scene so cuts between cleaned regions share one continuous bed. In Mellius you can request dialogue cleanup conversationally and then adjust per-clip intensity on the timeline, with undo protecting every step.
- Classify the noise before choosing a tool or preset.
- Reduce in light passes and re-check sibilants each time.
- Edit out unusable audio instead of filtering it harder.
- Lay continuous room tone under cleaned edits.
Rebuild presence and finish the mix
Noise removal often takes warmth and air with it, leaving dialogue thin. Restore body with a small low-mid lift, tame harshness with a narrow cut where the processing rings, and add a touch of consistent ambience so the voice sits in a space instead of a vacuum. Then set levels: dialogue anchors the mix, music ducks 12 to 20 dB beneath speech with per-phrase keyframes, and effects punctuate without masking consonants.
Validate on imperfect speakers. Phone speakers expose harshness, laptop speakers expose thinness, and earbuds expose gating artifacts — check all three before delivery. Generate captions from the cleaned track and confirm timing survived the edit, keeping text clear of faces on vertical crops. Common mistake: soloing the dialogue, perfecting it in isolation, and discovering it fights the music the moment the full mix plays.
Compare the workflows
| Approach | Best fit | Trade-off to consider |
|---|---|---|
| AI suppression passes | Steady hiss, hum, and room tone | Overuse creates metallic artifacts |
| Manual EQ and gating | Surgical problems and tonal shaping | Slower and demands more skill |
| Re-record or replace | Wind damage and overlapping voices | Costs session time or ADR effort |
A practical sequence
Put it into practice
- 01
Edit, then reduce lightly
Remove unusable audio first, then apply noise reduction in small staged passes.
- 02
Restore and bed
Rebuild warmth with gentle EQ and lay room tone under every cleaned cut.
- 03
Mix and verify
Duck music under dialogue, caption the result, and check on phone, laptop, and earbuds.
Frequently asked questions
Why does my dialogue sound robotic after noise removal?
The reduction was pushed past what the signal supports, stripping harmonics along with the noise. Dial back the last pass, restore some low-mid body with EQ, and accept a faint noise floor. A quiet human voice always beats a silent synthetic one.
Can AI noise removal fix wind or clipping?
Only mild cases. Heavy wind buffets and clipped peaks destroy the underlying waveform, so suppression just exposes the damage. For those sections, cut around the damage, use alternate takes, or re-record the line.
Should noise removal come before or after the music mix?
Before. Clean the dialogue first so you know what the voice actually sounds like, then fit music and effects around it with ducking. Mixing first forces the cleanup tool to fight your music bed.
Continue reading
Primary sources
Product capabilities change. These official references support the comparisons above.
Describe the edit. Keep the timeline.
Mellius turns conversational direction into cuts, captions, graphics, motion, color, and sound you can still inspect and change.
Start editing free