Mastering AI music: Suno, Udio and loudness

🎚️ Generating and mastering are two jobs

The generator composes and plays the song. It doesn't think about the final preparation for real players. It's like a phone photo: the composition is great, but before you post it you still adjust the contrast and sharpness.

What mastering does

It balances the frequencies so nothing stings and nothing is muddy.
It looks after the dynamics so the quiet parts are heard and the loud ones don't crackle.
It sets the loudness for the platform the song is going to.

Why even more so with AI

AI doesn't record sound, it guesses it, and compresses it along the way. Compression always throws something away.
A prompt won't fix it. “Studio quality” in the style doesn't raise the quality.
AI songs tend to have typical flaws that a microphone recording doesn't.

🔍 Typical flaws of AI sound

Metallic highs

A glassy, hissing sheen, mainly on vocals and cymbals. The generator guesses detail it doesn't have.

Fog in the mids

Instruments blend together and lose their outline, mainly in dense arrangements.

The bass drifts to the sides

The bass should play in the middle. When it drifts to the sides, it sounds odd on headphones and gets lost on a big speaker.

Stereo can be split into the mid (what's the same in both speakers) and the sides (the difference between them). In AI songs, flaws often hide in the sides. Cleaning up the sides and pulling the bass into the middle are among the most effective steps.

📢 The loudness trap

The song is quieter than the hits on Spotify, so I push it up with a limiter. But that breaks more than it fixes.

After loudness normalisation both play equally loud. But the overpushed master has lost its dynamics and its flaws are louder.

Why it doesn't work

An AI song is often heavily compressed straight out of generation. Another limiter squeezes out what's left of the dynamics.
The flaws get louder along with the music. They don't disappear, they're just louder.
Platforms normalise loudness anyway. Spotify plays at −14 LUFS, so a louder master doesn't sound louder.

The right order

First clean up the sound: highs, mids, bass into the middle.
Then loudness around −14 LUFS and the peak (true peak) below −1 dB. For a louder master, below −2 dB, as Spotify advises.

✅ What mastering can and can't do

It can

Soften metallic highs and harsh sibilance.
Reduce the fog in the mids.
Pull the bass into the middle so it holds up on all speakers.
Set the loudness for the platform and balance the sound for headphones, the car and the phone.

It can't

Bring back detail the AI threw away during generation. It's like enlarging a small photo: no new pixels appear.
Separate instruments that are blended in one mix.
Fix an overcrowded arrangement. A new generation with fewer instruments helps there, see Suno: arrangement.

⚡ What you can do right now for free

Without tools

Don't chase loudness with a limiter. The platform normalises it anyway.
Listen to the song on three systems: headphones, the car, a cheap Bluetooth speaker. Each flaw shows up somewhere else.
Download in WAV, not MP3. More compression makes the sound worse.

With a tool

Use a tool that's built with AI sound in mind. A generic “one click” often just adds loudness and leaves the sides and the bass alone.
For mastering on my own computer I built PD2AM. It's also handy as a step before making a video in PD2V.

📋 Prompts to copy

For ChatGPT, Claude, Gemini or Grok. A chat AI usually can't hear your song, so give it the numbers from a loudness meter or describe what you hear.

Explain my track's measurements ChatGPTClaudeGeminiGrok

I measured my song from [Suno / Udio], genre [genre]. The meter values:
- integrated loudness: [e.g. −9.5 LUFS]
- true peak: [e.g. +0.3 dBTP]
- loudness range (LRA): [e.g. 3.1 LU]

The song is going to [Spotify / YouTube / both].
Explain simply what these numbers mean, what's outside the platform's recommendations and what I should change in the mastering. For every piece of advice, say why. If something can't be told from the numbers, say so.

I'll describe what I hear ChatGPTClaudeGeminiGrok

I have an AI song ([genre]) and this is what bothers me about it:
[e.g. the bass booms in the car, the vocals hiss, it's oddly wide on headphones, the chorus blurs together]

Tell me which typical flaws of AI sound these most likely are and what to do about them in mastering. Use plain words and order the steps from the most important.
When something can't be fixed by mastering and only a new generation helps (for example with fewer instruments), say so straight away.

A check before release ChatGPTClaudeGeminiGrok

I'm about to release an AI song on [Spotify via DistroKid / YouTube / both].
Give me a short checklist: file format and quality, loudness and true peak, listening on several devices, title and description, and what to check in the terms of the generator ([Suno / Udio]) and the distributor regarding AI music.
For things that change, tell me to check them at the source, and don't give numbers you're not sure about.

← Suno: a better song The whole music section