AI mastering can be good enough for a release, and it can also miss. You find out for your track by comparing the AI master with your mix, and with any other master you have, at matched loudness and then blind. Turn every file down to the quietest one by integrated LUFS, switch between them in sync on the same sections, check true peak and tonal balance, and run a blind preference test on the closest two. If you pick the AI master most of the time without knowing which is which, it is the better master for that song.
The general method for any set of masters is in the guide on how to compare masters. This guide covers what changes when one of them comes from an automated service.
Is AI mastering good?
There is no answer that holds for every track. The same service, on the same setting, can suit one mix and not the next. A mix with a balanced low end and controlled peaks gives any mastering process less to fix. A mix with problems may need a decision a preset doesn't make.
Most of the arguments you read about it come from comparisons that weren't fair. A master is almost always louder than the mix, and ears are less sensitive to lows and highs at low levels (ISO 226, the Fletcher-Munson curves). So a louder file sounds fuller and brighter, and a difference of about 1 dB can bias the choice. An AI master played 8 dB louder than the mix will sound like an improvement whether it is one or not. A comparison where you know which file is the AI master is biased too, in whichever direction you expect.
Remove both biases and one question is left. Which file do you prefer when they play at the same level and you don't know which is which?
Set up a fair test
- Use the same mix file for every master: the AI service, a second service or preset, and a mastering engineer if you have one. Then the mastering is the only thing that differs.
- Download the full-quality file each service delivers, such as a WAV, and compare that. A preview that plays in the browser may be processed or encoded differently from the file you'd release.
- If the service lets you choose a style, intensity or loudness, treat each setting as a separate master. Pick the two or three you'd actually consider.
- Include the unmastered mix. Every master should sound at least as good as the mix at matched loudness.
- Rename the files neutrally, for example M1, M2 and M3, so the names don't tell you which is which during the blind part.
Match loudness first
Measure each file's integrated LUFS and turn every file down to the quietest one. Don't turn the quiet one up, or it may clip. Here is a typical set before and after matching.
| File | Integrated | Turn down by |
|---|---|---|
| Mix | −18 LUFS | 0 dB |
| AI master, default setting | −8 LUFS | 10 dB |
| AI master, gentle setting | −11 LUFS | 7 dB |
| Engineer's master | −10 LUFS | 8 dB |
Unmatched, the default AI master is 10 dB louder than the mix and 2 dB louder than the engineer's master, so it starts with an advantage that has nothing to do with quality. On Spotify that advantage disappears anyway. Spotify plays music at −14 LUFS by default and turns the −8 LUFS master down 6 dB and the −10 LUFS master down 4 dB, so both play at the same level. The level matching guide explains the method in more detail.
In ABSeq, load the mix and up to three masters into A, B, C and D. Auto Gain measures each file and turns every track down to the quietest one, and the A/B key switches to the next track at the same moment in the song.
What to listen for
Switch between the mix and each master on the same three sections: the first verse, the loudest chorus and the quietest part. Listen for one thing at a time.
- Low end. Is the kick and bass heavier or thinner than in the mix, and does the kick keep its attack?
- The loudest chorus. Does the level pump or the mix flatten when everything plays at once?
- Vocal and upper mids. Is the vocal harsher or more sibilant than in the mix?
- Quiet parts. Has the master raised them so far that the verse and chorus feel the same size?
- Width and depth. Does the stereo image or the reverb tail change in a way you didn't ask for?
Then check the numbers. True peak should sit at or below −1 dBTP, and at or below −2 dBTP for masters louder than −14 LUFS, which is Spotify's recommendation. A −8 LUFS master at −0.2 dBTP misses it, and inter-sample peaks can clip when a service encodes it to AAC, MP3 or Ogg. If the service offers a ceiling setting, set it lower and render again.
For tonal balance, commercial mixes tend to follow a slope of about −4.5 dB per octave. That is a general approximation, not a genre target, so use it to spot a master that leans far from your mix and references rather than to grade it. The tonal balance guide shows how to read the curve. An A/B difference graph goes one step further and shows, band by band after loudness matching, where one master has more than another, so "M2 sounds boomy" becomes "M2 has 2 dB more at 63 Hz than the mix". Confirm small low-end differences by ear before you send notes or switch settings.
Test the closest two blind
Drop any master that fails the checks above. Then compare the two closest in a blind preference test. Each round plays the same short passage from both, in shuffled order, and you pick the one you prefer. You see which file was which only at the end.
Run at least 10 rounds. Picking the same master 9 or more times in 10 is unlikely to happen by chance (p ≈ 0.011). If your picks split about evenly, you can't reliably prefer either one, and either will do for that song. The blind A/B test guide covers passage length and how many rounds to run.
ABSeq's Blind Test runs this on tracks A and B, with Auto Gain on by default. You set 1 to 50 rounds and the seconds per passage, and the file names appear only in the results. It is a preference test, so it tells you which master you like more, not whether you can tell them apart.
AI mastering vs a mastering engineer
The comparison method is the same for both. What differs is how you get a second version.
| AI or online service | Mastering engineer | |
|---|---|---|
| How you ask for a change | Change a style, intensity or loudness setting and render again | Send notes with a time, a part and a number |
| What you compare | Each new render against the last one and the mix, matched | Each new round against the last one and the mix, matched |
| What to check | Loudness, true peak, tonal balance, the loudest and quietest sections | The same |
If the engineer's master wins the blind test, you have your answer for that song. If the AI master wins, or your picks split evenly, the AI master is good enough for that song, and you can decide on price, turnaround or how you prefer to work. Run the test again on the next release. The result for one track doesn't carry over to a different mix.
Questions
Is AI mastering good?
It depends on the mix and the setting, so test it on your track. Match the AI master and your mix to the same integrated loudness, compare them on the same sections, check true peak and tonal balance, and run a blind test against any other master you have. If you prefer the AI master blind, it is good enough for that song.
How do I do an AI mastering comparison?
Master the same mix file with each service or setting, download the full-quality files, and turn them all down to the quietest one by integrated LUFS. Switch between them in sync on a verse, the loudest chorus and the quietest part, drop any that fail true peak or tonal balance checks, and test the closest two blind.
Is AI mastering better than a mastering engineer?
Neither wins by default. Compare the two masters of the same mix at matched loudness without knowing which is which. If you pick one 9 or more times in 10 rounds, prefer it for that song. If your picks split evenly, both work, and you can choose on price, turnaround or how you like to give feedback.
Is mastering with AI good enough for Spotify?
This is what people mean by searches such as "masterización con IA" and "masterização com IA". Spotify plays music at −14 LUFS by default and turns louder masters down, so an AI master's loudness gives it no advantage there. Check that its true peak is at or below −1 dBTP, or −2 dBTP if it is louder than −14 LUFS, and judge the sound at matched loudness.
Why does the AI master sound better than my mix?
Often because it is louder. A master is usually several dB louder than the mix, and a louder file sounds fuller and brighter. Turn the master down to the mix's integrated loudness and compare again. If it still sounds better, the mastering helped.