Clean up bad audio with AI before you re-record
What the enhancement tools genuinely fix, and the one problem they cannot.
Before you start
- A problem recording
- Any audio editor
What you will be able to do
- Identify which of the four common audio problems you have
- Apply repair in the order that does not fight itself
- Recognise the one failure that means re-recording
A recording comes back with a hum, a room, or a neighbour's lawnmower, and the assumption is that it has to be done again.
Often it does not. But the tools are strong on some problems and hopeless on others, and the difference is not obvious until you have wasted time on the wrong one.
Diagnose before you process
Hum, hiss, room and interruption are four different problems with four different fixes.
Listen once and name it. Hum is a steady low tone from electrics. Hiss is broadband noise from gain. Room is reverb — the recording sounds distant and boxy. Interruptions are discrete events: a door, a cough, a car.
Running a general "enhance" over all of them at once is what produces that underwater, over-processed voice, because each tool is trying to solve a different problem simultaneously.
Take the noise out first, gently
Noise reduction before anything else, and at half the strength you want.
Modern speech-isolation models are strikingly good at hum and hiss. Apply reduction first, because everything downstream — de-reverb especially — works better on a clean signal.
Use less than the maximum. Aggressive reduction removes the breath and sibilance that make a voice sound like a person, and that damage cannot be undone later.
- A/B against the original constantly. Ears adapt to processed audio within about a minute and stop hearing the artefacts.
De-reverb, and accept a partial win
Room is the hardest of the four and the tools have genuinely improved.
Speech models can now pull a voice noticeably forward out of a room, which was not really possible a few years ago. It is still the least complete fix of the four.
Expect improvement rather than resolution, and stop early — pushed hard, de-reverb produces a thin, phasey voice that sounds worse than the room did.
Warning Know the case that cannot be repaired
Clipping is missing information. Nothing recovers it.
If the recording was too loud at the source, the peaks were flattened and that data does not exist anywhere in the file. De-clipping tools reconstruct a plausible curve and it sounds like a reconstruction.
Light clipping on a few peaks is survivable. Sustained clipping through a whole take is a re-record, and recognising it early is worth more than any plugin.
Try repair before booking a re-record — it takes ten minutes. Just stop the moment you hit the clipping case, because that one does not have an answer.
Common questions
Was this guide useful?
88% of readers found this useful
Read next
Turn a script into a watchable AI video
Generating clips is the easy part. What separates a watchable AI video from a slideshow is pacing, continuity and audio — none of…
Clone a voice responsibly and legally
Voice cloning has real uses: accessibility, localisation, and your own voice at scale. All of them depend on getting consent and d…
Connect two apps with an AI step in the middle
The genuinely useful automations are not the clever ones. They are a trigger, one AI step that makes a small judgement, and a writ…
Choose your first AI assistant without overthinking it
Every comparison table lists twenty differences and only three of them change your day. Here is how to pick in ten minutes and get…
