What mastering actually is
Mastering is the final step of production: processing applied to the finished stereo file — not to the individual instruments inside it (iZotope's overview is a good neutral reference). The master engineer — human or algorithmic — receives one wave file and shapes it as a whole: overall loudness, overall tonal balance, peak safety, consistency. That "as a whole" is the key phrase in the job description, because it defines both the power and the limits of the step. Everything mastering does, it does to the entire track at once.
What mastering reliably fixes
Your track is quiet next to commercial releases. This is mastering's home turf. Raising perceived loudness to a competitive level without audible damage is the problem the whole discipline grew up around, and it's the single most common issue with raw AI generations.
The overall tone is a bit dull, dark, or thin. Broad-stroke tonal balance — the whole track needs more air on top, less boom at the bottom — is a whole-file adjustment, which is exactly the kind mastering can make.
Peaks that would clip or distort after streaming encoding. Controlling true peaks and leaving sensible headroom is a mastering deliverable, and it's what keeps a master from crackling after platforms convert it to lossy formats.
Tracks on one release that don't sit together. Running every track through the same mastering pass gives an EP a shared character and level — we wrote a whole guide to release consistency about this.
What mastering can't fix
A buried vocal — or any balance problem between elements. Mastering works on the summed file. There is no "vocal fader" anymore: boosting the vocal's frequency range boosts every instrument living in that range too. The industry has repeated this for decades about human-made music — "mastering can't fix a bad mix" (one of many mastering engineers saying so) — and it's just as true when the mix came out of a model.
Artifacts baked into the generation. Garbled consonants, watery textures, cymbals that fizz instead of ring — these are part of the audio itself. Mastering makes the file louder and clearer, which means it makes the artifacts louder and clearer too. A polish step cannot subtract what's printed into the sound.
Arrangement mud. If the low end feels crowded because bass, kick, and pad all occupy the same space, that's a decision inside the music. Mastering can tilt the overall balance, but it can't un-stack instruments.
Songwriting and structure. The chorus that arrives too late, the outro that overstays — no processor touches these.
A low-quality source file. Mastering amplifies what's in the file, compression artifacts included. Starting from the best available export (WAV over MP3) matters more than any setting downstream.
One more thing mastering can't do: win by sheer volume. Streaming platforms normalize loudness on playback — master dramatically louder than everyone else and Spotify simply turns it down (Spotify's guidance). What survives normalization is balance and punch, not the number you hit.
"Go back and fix the mix" — but there is no mix
Here's where the classic advice breaks for AI music. Every traditional article ends the same way: reopen the session, rebalance, re-export. A Suno or Udio creator has no session. You have a finished stereo file and a prompt box. That doesn't mean you're stuck — it means your "mixing stage" lives somewhere else: in generation.
Re-generate. The fastest fix for a buried vocal or a harsh generation is another take. Generations are cheap; repair time isn't. If a track fights the master, it will usually keep fighting — a cleaner generation five minutes later beats an hour of processing.
Steer with the prompt. Style and production tags ("clear lead vocal," "minimal arrangement," "acoustic") shift the balance the model prints. It's imprecise, but it's the closest thing to a fader you have.
Use stems where offered. Some platforms can export separated stems. Quality varies, but for a genuinely buried vocal, rebalancing two stems — even roughly — fixes what no full-mix process can.
Then master. Once the generation is right, mastering does its actual job: competitive loudness, tonal polish, safe peaks, consistency. In that order — generation first, mastering last — each step fixes what it's built to fix.
A 30-second diagnosis
Before re-mastering a track for the third time, name the symptom. Quiet next to commercial tracks; dull or thin overall; distorts after upload; tracks don't sit together as a release — mastering problems; process (or re-process) the file. Can't hear the vocal; garbled or fizzy details; muddy low end; the song drags — generation problems; no master will fix them, so go back to the prompt. If you're not sure which side a symptom is on, our ear-only checklist walks through the listening tests that tell you.
Why a mastering app is telling you this
We build Tonekai, a local AI mastering app for macOS — so it might seem odd to publish a list of things our product category can't do. But overpromising is how creators end up re-mastering one flawed generation ten times. Tonekai is built for the honest version of the workflow: drop in a finished track and it handles analysis, loudness, tonal balance, and peak control automatically; the built-in A/B lets you hear exactly what changed — and, just as usefully, what didn't, which is your cue that the problem lives in the generation, not the master. Everything runs on-device (no uploads, Apple Silicon native, macOS 14+), so the re-generate → re-master loop takes minutes. Free to try, with a one-time Pro Unlock — no subscription — for WAV/AAC export and unlimited mastering.
Download Tonekai on the Mac App Store — free to try.
Frequently asked questions
My track sounds better mastered but still not "professional." Is the master bad? Probably not. Level-match the loudness and compare your original against a commercial reference. If the gap is already there before mastering, the gap is in the generation or arrangement — mastering only carried it forward.
Can EQ on the master bring the vocal forward at all? A little, sometimes — a gentle lift in the presence region can add a sense of clarity. But everything sharing those frequencies comes up too. Treat it as a nudge, not a fix.
Should I skip mastering until I get a perfect generation? No — work in the right order rather than skipping steps. Get a generation that survives a critical listen, then master it. An unmastered good take beats a heavily mastered bad one, but a mastered good take beats both.
Do these limits apply to human-made music too? Almost all of them, yes. The difference is the remedy: a producer reopens the mix; an AI creator re-generates. The mastering stage itself has the same job description either way.
The bottom line
Mastering is the last 10% — loudness, tonal polish, peak safety, consistency — and done well, that 10% is the difference between a demo and a release. But it processes the whole file as one thing, so it can't rebalance, repair, or rewrite what's inside. When a mastered track still sounds wrong, don't reach for a fourth mastering pass. Name the symptom, fix it at the layer it lives on — and for AI music, that layer is usually the prompt box, not the processor.