AI Music
How to Master AI Music in 2026: Four Routes to a Release-Ready Track

Your track sounds finished in the generator and then sits half a step behind everything else in the playlist. The usual advice is to master AI music to a platform loudness target, and the usual page gives you a table of targets for six platforms. We went and read what those six actually publish. One of them publishes a number.
So this is the version of the mastering guide that separates the documented from the repeated, and then walks the four routes people are actually using in 2026 — Suno’s own Remaster button, an LLM with a Python sandbox, a free meter you run yourself, and paying somebody. Every figure below was read at its source on 21 October 2026, and where a source says nothing, this page says that instead of filling the gap.
The short version
Spotify is the only major platform that publishes a mastering target — -14 dB integrated LUFS, true peak below -1 dBTP. Apple’s own guide has no LUFS figure in it at all; TIDAL’s and Deezer’s help centres return nothing for “loudness”.
Suno’s Remaster re-generates the song instead of mastering the file, so it is not route one for a track you already like.
Mastering does not change how a platform labels your release. YouTube’s own guidance puts automated AI mastering under “No Gen AI” — the generated audio underneath is what gets declared.
Why an AI export is not a master
A generator is solving for a convincing two-and-a-half minutes, not for a delivery spec. The file it hands back is a finished-sounding mix with none of the decisions a mastering pass exists to make, and the symptoms are consistent enough to list.

Loudness is unpredictable from render to render, so an album assembled out of twelve generations steps up and down between tracks. Headroom is wherever the model left it, which matters because lossy encoding adds peaks that were not in your file. And the balance is a statistical average of the training material rather than a choice — loud in the middle, soft at the edges, with transients already squashed by whatever the model learned a mastered record sounds like.
That last one is why mastering an AI track is not the same job as mastering a mix. You are not adding the final 10% to something that was built with headroom; you are working on a file that has already been through something that behaves like a limiter, with no stems behind it unless you paid for them. Keep the goal modest: in spec, consistent, and undamaged.
Which loudness numbers are actually published
Every mastering guide in the search results carries the same table: Spotify -14, Apple -16, YouTube -14, Amazon -14, TIDAL -14, Deezer -15. We tried to source each cell from the platform itself. Here is what came back.
| Platform | Published for people making records? | What the source actually says | Checked |
|---|---|---|---|
| Spotify | Yes | Spotify for Artists: target “-14dB integrated LUFS” and keep true peak “below -1dB TP”, or below -2dB if the master is louder than -14. | 21 Oct 2026 |
| Apple Music | No | The Apple Digital Masters guide contains no LUFS or LKFS value. It asks you to “leave at least 1 dB of headroom” and to master so the track “captures your intended sound, regardless of playback volume”. Sound Check sets level at playback, per track or per album. | 21 Oct 2026 |
| YouTube / YouTube Music | No | No loudness specification for music uploads in the help centre; playback is normalised, the figure is not documented for creators. | 21 Oct 2026 |
| TIDAL | No | Its help centre search returns zero articles for “loudness” and zero for “normalization”. | 21 Oct 2026 |
| Deezer | No | Zero articles for “loudness” in its help centre. The -15 figure circulating online traces to a forum post, not a specification. | 21 Oct 2026 |
| Amazon Music | No | No published mastering target found in Amazon Music for Artists. | 21 Oct 2026 |

The difference matters more than it looks. A published target is a submission guideline you can master to; an undocumented one is a measurement of what a player did on the day somebody measured it. Those numbers are not useless — they are roughly right, and we keep the working set on our LUFS targets for Spotify and streaming page — but you should know which one you are obeying. Both ends of the table are one click away: read Spotify’s loudness normalisation page and then search Apple’s own Apple Digital Masters guide for the string “LUFS”. It is not in there.

What normalisation does, and what it does not
Normalisation is a volume change, applied at playback, to the whole track. It is not processing. Nothing is compressed, nothing is re-balanced, and your file is not modified — the player simply decides how much gain to put in front of it.
Which means the thing people fear — “Spotify will crush my master” — is not what happens. What happens is subtler and worse: a track squashed to -7 LUFS gets turned down seven or eight dB to sit beside everything else, and arrives with all of its dynamic range already spent and none of the loudness it paid for. The master that was made quieter and cleaner arrives at the same level with its transients intact. That is the whole argument against chasing loudness on a streaming release, and it is why the limiter is the last place to look for impact.
The one asymmetry to know: platforms are cautious about turning quiet tracks up, because gain applied to a file that already peaks near zero would clip it. Spotify documents exactly that behaviour — a quiet track with little headroom gets only as much gain as its peaks allow. So a master at -20 LUFS does not reliably arrive at -14. Leaving real headroom is what lets the platform do its job.

Four routes to a mastered AI track
Once you accept the generator will not do it, there are four things people actually do, and they are not interchangeable. They differ in what they change about the file and in what they can never change.
| Route | What it changes | What it cannot do | What to check afterwards |
|---|---|---|---|
| Suno Remaster | Nothing about your file — it generates a new render of the song on a current model | Preserve the take you liked; raise the loudness of an existing export | Whether the new render is the same song you wanted |
| ChatGPT / an LLM sandbox | Real DSP: gain, EQ, compression, limiting, loudness measurement, written as code and run on the file | Hear the result; make a judgement call; know whether it helped | The measured integrated loudness and true peak of the file it hands back |
| A meter plus ffmpeg or a DAW | Exactly what you tell it to, with a number on the before and the after | Make taste decisions for you | Nothing — the measurement is the output |
| A service, or a person | A finished master; a person also makes judgements about the record | Fix the arrangement or the performance underneath | That the return is in spec and that you can still hear the top end |
Notice what the column headings do to the question. “Which is best” is unanswerable without the track in front of you. “Which one can do the thing I need” usually has one answer, and for a track you already like it is never the first row.
Route 1: Suno’s Remaster button
The name is the problem. In audio, remastering means going back to the source and preparing it again. Suno’s Remaster re-generates the song on a current model — it is a new performance of the same prompt and lyrics, not a new treatment of the file you downloaded. If what you liked was a particular vocal phrase or a particular drum fill, there is no guarantee it survives.
It is also metered. A remaster is a generation and a download, and downloads are the thing Suno counts. We went through the mechanics, the costs and the model-retirement catch in Suno Remaster: what it actually does to an old song, and the short version is that it is a good tool for the problem it solves — an old track rendered on a model that no longer exists — and the wrong tool for the problem people reach for it with, which is loudness.
The editing environment is no help here either: the reporting on Suno Studio 2.0 counts seven effects — compressor, convolution, delay, distortion, EQ, gate, reverb — with no limiter and no loudness meter. You cannot master inside a room with no meter in it.
Route 2: handing it to ChatGPT
This is the route that has arrived in the last few weeks, and it is worth being precise about, because two completely different things get described with the same sentence.
| What the person does | What actually happens | Where it goes wrong |
|---|---|---|
| Describes the track and asks for mastering settings | The model writes a plausible chain — EQ moves, compressor ratios, a limiter ceiling — from the description alone | It has not heard the track. The advice is a genre average, and it will be stated with the same confidence whether or not it fits |
| Uploads the file and asks for it to be mastered | Code runs in a sandbox: the audio is read as samples, filters and gain are applied, a new file comes back. This is real processing | The model still cannot hear the output. It reports success from the code completing, not from the result sounding right |
| Uploads the file and asks for it to be measured | A loudness and true-peak measurement, which is arithmetic and reliable | Very little — this is the use that plays to the tool’s strengths |
The second row is the one people argue about, and both sides are half right. It genuinely processes the file — a gain stage written in Python is as real as a gain stage in a DAW. But every judgement in mastering is a listening judgement, and that is the one faculty the sandbox does not have. The honest framing is that you are using a very fast assistant who writes competent DSP code and then describes the result from the log rather than the speakers.
Which makes it usable, with a rule: never accept a described outcome, only a measured one. Ask for the integrated loudness and true peak of the input, the same two numbers for the output, and listen to both files yourself. If those numbers are not in the reply, you have been told a story about your track.
Route 3: the free route you can check yourself
Everything in the previous section is available directly, for nothing, with the measurement built in. ffmpeg implements the EBU R128 loudness standard in its loudnorm filter, which both measures and normalises. Measuring first is the step people skip:
# 1. Measure what you actually have
ffmpeg -i song.wav -af loudnorm=I=-14:TP=-1:LRA=11:print_format=summary -f null -
# 2. Normalise to the one published target, writing a new file
ffmpeg -i song.wav -af loudnorm=I=-14:TP=-1:LRA=11 -ar 48000 song-master.wavThe first command prints input integrated loudness, input true peak and loudness range, then the same figures for what the output would be. That is the before-and-after nobody in the tool roundups publishes, and it takes about two seconds. If you would rather not touch a terminal, a free loudness meter plugin in any DAW gives you the same reading.
What this route does not give you is tone. Loudness normalisation makes a quiet track the right loudness; it does not fix a dull top end or a boxy low-mid, which is the actual complaint behind most “my AI track sounds flat” posts. For that you need an EQ decision, which means either learning to make one or paying someone who already has. Run the measurement anyway: knowing you are at -9 LUFS with -0.2 dBTP changes what you ask for next.
Route 4: a service, or a human
Automated mastering services do the thing route three does plus the tone decision, by matching your track against a model of finished records. They are quick, consistent and genuinely fine for a single-track release. We are not printing a price table here, because every vendor’s pricing moves and we could not read a single one of their pricing pages from source this week — check the number on the vendor’s own page rather than in anyone’s roundup, including the roundups that rank them.
A human engineer is the other end, and the live question in the forums is not price but disclosure: do you tell them the track came out of a generator? Tell them. An engineer prices by how much work a file needs, and a generated master is an unusual job — no stems unless you bought them, reverb baked into the vocal, and transients already shaved. Some engineers decline AI work on principle, and finding that out in the first message costs nothing, while finding it out after a deposit costs the deposit.
If the complaints are about the record rather than the file — a vocal that never commits, drums with no room — mastering is the wrong department entirely. That is a rebuild, and we walk through pulling stems into a DAW in why labels reject Suno tracks.
Does mastering change how a platform labels you?
This became a real question on 14 October 2026, when YouTube Music began labelling AI-generated songs. It is worth answering precisely, because the fear — that running a track through an AI mastering tool is itself an act of AI generation that gets you flagged — is backwards. YouTube’s own guidance on Gen AI designations answers it with worked examples.
| Step | How YouTube's guidance treats it | Who decides |
|---|---|---|
| Prompting a generator and downloading the result | Fully Gen AI | Your distributor’s declaration on delivery |
| Automated AI mastering of the finished track | No Gen AI | Named in YouTube’s own worked examples |
| Standard pitch correction on a sung vocal | No Gen AI | Named in YouTube’s own worked examples |
| Delivering with no declaration at all | YouTube may designate it for you | YouTube’s own detection signals |
Read the second and third rows together with the first and the logic is clear: the designation is about how the content was created, not how it was finished. Mastering — by hand, by a service, or by an AI mastering tool — is post-production, and YouTube says so in as many words. The generated audio underneath is what gets declared, and the declaration travels in the delivery metadata from your distributor.
Which is the boundary to hold on to, because plenty of marketing implies otherwise: no audio processing changes a label that is driven by a declaration. Not mastering, not watermark removal, not re-rendering. Those steps address screening and artifacts, which is a different gate with a different owner — we separate the two on distributing AI music without getting flagged.
The order of operations before you upload
Mastering sits in the middle of a sequence, and running the steps out of order is how people end up mastering twice. Clean first, because anything you strip or repair afterwards changes the level you just set.
WAV over MP3 where your plan allows it — more to work with, and no encoder artifacts stacked under the ones already there.
Artifact repair and marker removal change the signal. Do them while there is still headroom, not after limiting.
The one published target. Hitting it keeps every other platform's normalisation out of your way.
Print the integrated loudness and true peak of the finished file. A master you have not measured is a guess.
Answer your distributor's AI questions accurately. That field, not your processing chain, is what sets the label.

Step two is the one unique to an AI release, and it is where Undetectr fits: it is built for the artifacts and the embedded markers that get a generated file bounced by automated screening at a distributor, and it sets level in the same pass. To be exact about what that is and is not — it addresses screening and audible artifacts. It does not change a platform’s AI label, because nothing does; that comes from the declaration, as above.
Our recommended tool
Clear the screening step, then master
Undetectr targets the generation artifacts and embedded markers that get AI tracks bounced by automated distributor screening, and sets level in the same pass. It has no effect on how a platform labels your release.
And when the master is finished
Worth saying plainly, because a mastering guide can imply that loudness is what stands between a track and an audience. It is not. Mastering is finishing work — it removes a reason to skip, it does not create a reason to listen, and the hard problem in AI music remains that nobody can find your release. The routes that do not depend on being discovered are worth more than another dB: a paid sync placement is a person reading a brief and choosing a cue, and selling direct to the listeners you already have keeps all of the money rather than a per-stream fraction of it. A finished master helps with both, which is the honest reason to do it properly.
Related reading on Erasy
- LUFS for Spotify and streaming: the working numbers
- Suno Remaster: what it actually does to an old song
- Cleaning AI-generated music for release: the checklist, in order
- Suno v6 sounds muffled: what changed, and what fixes it
- Distributing AI music without getting flagged
- 1,000 streams versus selling direct: the arithmetic
Frequently asked questions

Eddie Mathews — Editor, Erasy
Eddie reads the help-centre article and the specification rather than the roundup, and marks the difference between what a company has documented and what a tutorial has assumed. Where the public record does not settle a question, this site says so.
Does AI music need mastering?+
What LUFS should I master AI music to?+
Is Apple Music really -16 LUFS?+
Can ChatGPT master my Suno track?+
Does Suno’s Remaster button master my song?+
Will mastering stop my track being labelled AI?+
Should I tell a mastering engineer the track was AI-generated?+
Read at source on 21 October 2026: Spotify for Artists on loudness normalisation; Apple’s Apple Digital Masters guide, searched in full for LUFS and LKFS, which do not appear in it; YouTube’s help-centre article on Gen AI designations, including its worked examples for automated AI mastering and pitch correction; the TIDAL and Deezer help-centre search APIs, which return no loudness articles; and the ffmpeg filter documentation for loudnorm, whose two-pass output we ran in order to quote what it prints. The count of effects in Suno Studio 2.0 is AI Musicpreneur’s reporting of 10 September 2026, not our own test of the product; that article discloses a paid partnership with one of the tools it ranks. We have not measured a Suno export ourselves and no figure here is presented as one.