What to Fix Before Mastering an AI Track
2026-09-067 min
If a Suno or Udio track sounds harsh on the way to release, the problem is usually not what you did but the order you did it in. Codec residue gets louder along with everything else, so changing the order changes the result.
The short answer
- The order is clean first, master second. Reversed, you amplify the noise floor and then have to carve it back out.
- Make-up gain does not distinguish signal from noise. Push the track 6 dB louder and the hiss comes with it, 6 dB louder.
- An air-band boost is the worst offender. The hiss is strongest around 9-12 kHz, and the 10 kHz shelf everyone reaches for lifts exactly that.
- Most streaming platforms normalise loudness on playback, so pushing level harder does not make you louder — it just leaves you with more noise.
- Do not try to fix it by regenerating. Remaster and upscalers rebuild rather than restore, so the take you liked changes.
1. Why the order changes the result
Mastering does two broad things: it shapes tone and it makes the whole thing louder. The second one is the problem. Make-up gain after a compressor or limiter does not distinguish signal from noise. Make the track 6 dB louder and the codec hiss underneath it is 6 dB louder too.
It shows up worst in the quiet parts. Intros, breakdowns and fades are exactly where dynamic processing applies the most gain. On a recorded track there is room tone in those gaps; on a generated track there is a noise texture the codec put there. That is what gets amplified.
The air band makes it worse again. Mastering commonly adds a high shelf around 10 kHz for air, and in Honenuki's measurements the hiss in generated audio is strongest between 9 and 12 kHz. A move intended to add brightness lands directly on the hiss and lifts it.
So the order matters. Clean first and raise second, and what comes up when you raise it is the music.
2. The order to work in
Worked in this order, no step fights the next one. What to judge at each step is listed beside it.
| Step | What | How to judge it |
|---|---|---|
| 1 | Get stems if you can | Per-stem work beats working on the full mix. Vocals and drums need different amounts of attenuation |
| 2 | Remove the codec residue | Solo the residual and check that no music is in it. If there is, lower Strength |
| 3 | Put the texture back | Restore the high-frequency character that removal took with it. You do not need to end up brighter than the original |
| 4 | Mix | Levels, panning, EQ. From here it is ordinary mixing |
| 5 | Master | Now raising the level does not bring hiss up with it |
| 6 | Release | Platforms normalise loudness on playback, so a loudness war is mostly pointless |
Steps 2 and 3 happen together inside Honenuki. From step 4 on, use whatever you normally use.
Sometimes stems are not available. Then apply it to the full mix but start with Strength low — different stems need different amounts, and bundling them means working to whichever track is most sensitive.
3. Common mistakes
Even in the right order, these four come up constantly.
| Mistake | Why it fails | Do this instead |
|---|---|---|
| Burying it under loudness | Make-up gain raises the noise floor too, and platforms normalise playback anyway, so it does not even sound louder | Take it out first; decide loudness after the cleanup |
| Adding air for brightness | It lifts the range just above 9-12 kHz where the hiss is strongest | Add it after cleaning, and only as much as needed. Honenuki's Air exists to return what removal took |
| Leaning hard on a broadband denoiser | Cymbals and consonants go with it, and musical noise appears where there was none | Target narrow bands. Solo the residual and check whether music is in it |
| Regenerating with Remaster | That rebuilds rather than restores, so the take changes — and the output passes through the same codec and picks up the same residue | If you like the take, fix that file |
When you cannot tell whether you have gone too far, listen to the residual on its own. If you hear music in it you took too much; if you hear only hiss you are on track.
Questions
Before or after mastering?
Before. Codec residue and high-frequency hiss get louder along with everything else when you raise the level. Cleaning after mastering means carving out something that has already been amplified, and the music gets carved with it. Honenuki is not a mastering chain and applies no limiting or loudness normalisation.
How loud should I master?
Most streaming platforms normalise loudness on playback, so pushing harder does not make you louder — it just costs dynamics. Targets differ by platform and change over time, so check the current recommendation for wherever you are releasing. What matters here is what else gets louder on the way to that target.
What if I only have the full mix, no stems?
It still works. But different stems need different amounts of attenuation, so a single file means working to whichever element is most sensitive. Start with Strength low and raise it while listening to the residual.
Will cleaning get my release past screening?
We make no such claim. Honenuki removes audible codec artifacts and does not touch watermarks or provenance credentials. It has nothing to do with distributor or platform screening.
What if I only have an already-mastered file?
It still applies. But it has already been amplified, so there is more to take out, and a limiter has already flattened the transients, which leaves less room to restore. If you can get the pre-master file, use that instead.
Sign up and it is free during early access.
Sources and caveats
- The observation that hiss is strongest between 9 and 12 kHz, the processing band (6-14 kHz) and the attenuation caps (-40 dB and -20 dB) are Honenuki's own engine measurements.
- That make-up gain lifts the noise floor follows from the definition of dynamic range compression. It is not a claim that needs a citation.
- Loudness normalisation targets vary by platform and change over time. This article recommends no specific figure — check the current recommendation where you are releasing.
- This article makes no claim about passing release screening or evading detection.