Open a plugin you bought five years ago and one you bought last month. The older one waits for you to tweak something. The newer one has probably already had a look at your audio and formed an opinion.
That is the real story behind AI-powered audio plugins. Not the marketing badge, but a change in what the developers building them decided to make.
Here is what actually changed, which companies went which way, and what it cost some of them.
What Changed Inside AI-Powered Audio Plugins
A tour of the popular AI plugins in the game
A conventional plugin is maths you wrote down. An engineer decides how a compressor should behave, expresses that as code, and the plugin does exactly that every time. Predictable, and the same for everyone.
A model-based plugin works differently. Someone gathers a large set of audio examples, trains a model on them, and ships the trained result. The plugin no longer follows rules a person wrote. It follows patterns it found.
Two things follow from that, and both matter more than the badge on the box.
First, these tools analyze before they act. They listen to your material and propose a setting, rather than sitting at a default. Second, the behavior is shaped by the training data. A model taught on modern pop will hear your folk recording through that lens, and there is no knob for that.
None of which makes them worse. It makes them different in a way worth understanding before you hand one a mix.
Different Approaches
There was never one way to do this. All of the examples, and you can hear the difference.
iZotope: the assistant
Ozone’s Master Assistant builds a starting chain for you
iZotope is one of the first ones to set a standard for the industry to follow. Ozone‘s Master Assistant listens to your track, then builds a starting chain and hands it over. Every module stays editable afterwards.
That framing has stuck. The tool proposes, you dispose. Most assistive plugins still work that way today.
iZotope walks through what is new in Ozone 12
sonible: built this way from the start
smart:EQ 4 listens to a section, then proposes a curve
The Austrian company never had a conventional back catalogue to convert. Its whole line is built around analysis.
smart:EQ 4 carries 24 bands and a set of profiles, one universal, 21 for instruments and 22 for genres. It listens to a section of your track and generates a curve to fit the profile you chose. You can also load a reference track and have it build a profile from that. Its Group mode unmasks up to ten instances at once, which is the sort of job nobody enjoys doing by hand. MusicTech’s review of smart:EQ 4 goes deeper on how it behaves in a mix.
sonible demonstrates smart:EQ 4
Waves: a neural net pointed at one job
Clarity Vx: one narrow job, done with a neural network
Waves went narrow instead of broad. Clarity Vx does one thing, which is separating a voice from the noise around it, and the company is direct about the method. The product carries the line “Powered by Waves Neural Networks”.
This is worth noting because narrow problems suit models well. Voice or not voice is a question with a right answer. Taste is not.
Antares: a classic tool learns a new trick
Auto-Tune Vocal Compressor suggests settings you can still move
Antares has been shipping Auto-Tune since the nineties, which makes it the clearest case of a conventional company turning. Its Auto-Tune Vocal Compressor is billed as the first dual-stage vocal compressor with machine learning. It analyses the vocal, suggests compression settings, and lets you move them afterwards.
Antares walks through the Vocal Compressor styles
Underneath, it is still two compressors with four styles each. The model picks the starting point. MusicRadar covered the plugin and how its machine learning is used at launch.
What the Shift Cost
Five years of ownership changes behind the plugins on your drive
Building models is expensive. It needs data, research staff and time, and that kind of money usually arrives with investors attached. The last five years show what followed.
In 2021, iZotope and Native Instruments were grouped together under backing from Francisco Partners. The group took on Brainworx and Plugin Alliance and became Soundwide in 2022. Layoffs followed in January 2023, and by that June the whole thing had been folded back under the Native Instruments name.
Then it came apart. Native Instruments entered preliminary insolvency in January 2026. Francisco Partners sold the brand to InMusic in May, which we covered in our piece on the inMusic acquisition. In July, iZotope left again and was bought by Boris FX, a visual effects software company. Licences stayed active throughout. MusicTech has the full sequence of that break-up.
Accusonus went a different way and vanished. Meta bought the company in early 2022, and the plugins were sunset rather than carried on. People who had built sessions around them had to rebuild.
So the practical lesson has nothing to do with sound quality. A model-heavy plugin ties you to a company that needs serious funding to keep going, and that ownership can change more than once. Keep your sessions printable.
Generation at the Idea Stage, Hands After That
The Harmony Engine in Fluid Chords 2
Every plugin should answer this one question: Where should the tool decide, and where should you?
We build MIDI effects, and the Harmony Engine inside Fluid Chords 2 addresses this very clearly. Give it a key and the chords. It hands back you the chord transitions in a super effective way.
That is generation, and it is pointed at exactly the kind of narrow problem models handle well. You are asking for somewhere to start, not a finished record. A harmony suggestion costs nothing to reject.
What it hands back is the part that counts. Progressions arrive as MIDI. Move a voicing, drop a chord, keep two bars and bin the rest. Once it is in your session the plugin has no further opinion about it.
Everything downstream of that answers to your fingers, and that is a musical decision rather than a technical one. Strum, Flow and Offset shape how a chord bends. Real-Time Mode and Split & Play are performance controls. Expression is the one thing you cannot hand to a model and still call the result a performance.
Hear Fluid Chords 2 in action
How to Read the Claims
“AI-powered” covers everything from a genuine neural network to a lookup table with good marketing. Four questions sort it out quickly.
What does it hand back?
Editable settings, or a finished result? A tool that lands you on a chain you can pull apart teaches you something. One that returns processed audio and nothing else does not.
Is the problem narrow enough?
Separating voice from noise has a right answer. Choosing a vocal sound does not. Models do well on the first kind of problem and get opinionated on the second.
What was it trained on?
Few companies publish this, which is itself worth noting. Where a plugin offers genre profiles, check whether yours is there. If your music sits outside the training set, the suggestions will pull towards something else. Our piece on AI sampling versus sound libraries runs into the same issue from another angle.
Does it run offline?
A plugin that processes locally keeps working whatever happens to the company. One that calls a server does not. Given the ownership churn above, that is a fair thing to check before a tool becomes load-bearing.
Frequently Asked Questions
What makes a plugin AI-powered rather than just automatic?
The difference is where the behaviour comes from. An automatic plugin follows rules an engineer wrote. An AI-powered one follows patterns learned from training data. In practice, the second kind analyses your audio first and proposes a setting rather than sitting at a default.
Why did so many AI plugin companies get acquired?
Training models costs far more than writing DSP. That means outside investment, and investment eventually wants a return. iZotope passed through Francisco Partners, Soundwide, Native Instruments, InMusic and Boris FX inside five years. Accusonus was bought by Meta and its plugins were discontinued.
Will AI-powered audio plugins replace mixing engineers?
Nothing in the current generation suggests it. These tools analyse and suggest, and a person still decides what serves the song. We went into this properly in our look at AI mix assistants.
Does a smaller developer mean weaker AI features?
Not at all. It usually means narrower ones. Smaller teams tend to apply models to a specific job rather than across a whole range, and narrow problems are often where this technology works best anyway.