A lot of what these AI tools do would classically be considered trying to “fix” a subpar mix, rather than just mastering. This is often not a super transparent change to the sound. For example, right up front, the first thing Logic’s mastering assistant does is put a full spectrum, highly detailed EQ curve over the track, ostensibly to hit some genre-dependent targets. Then it brags to you about it by graphing out the curve. This is not something many professional mastering engineers would do. Their job often consists of taking all the creative decisions made up to this point in the process and make subtle changes that bring them all out, much like adding the right amount of salt to a dish when cooking or polishing a car that’s already been detailed. They need to do this while making the track loud enough to compete in the consumer market, without crushing the dynamics or adding uncomfortable harshness to the sound. Generally, I think these AI products need to do less, not more, to get better results.

That said, this over-zealous AI modification isn’t always a bad thing, as the democratization of music recording has ushered in an era where many more people of various aptitudes are mixing music! But it does mean that AI-based mastering solutions run the risk of ossifying the “sound” of music genres, and while they can gussy up a poor mix, they often clobber a mix that’s already great. I don’t know of many major label albums where any of the tracks were mastered with an AI-based solution, although Ozone is in extensive use in the industry for its non-AI features.

Disclaimer: I make a non-AI mastering audio plugin called Master Plan.

I did a music course last summer and met some prefessional musicians and also sound engineers working for big record labels. I saw specifically LANDR in action.

I think like all of these new AI tools, it can be a tool for you to use, but an amateur with untrained ears will still not make "professional-grade" masterings. But you can get maybe 80% there. If you're a hobbyist and don't know anything about mastering and can't afford to pay a professional, this is great!

Hmm.

> the system spits back a song that hits modern loudness standards and is punched up with additional clarity, EQ, stereo width, and dynamics processing

Translation: "AI" can now make all recordings indistinguishable from each other?

'Standards' in art are dangerous.

Just got done mastering a track (a soundtrack for video with voice over). Getting to the appropriate LUFS level with a limiter seems to me something that is easily automated, and I'm not sure you need AI to do it. But as usual, most of the moves I made with EQ, compression, saturation, and so on were creative choices where I was picking from among dozens of different possible tones and vibes (some in my head, some from reference tracks). I guess I don't doubt that AI can master a track if you say "make it sound like a Beyoncé record," and some of the automated tools give you vague choices about how you want it to sound.

But for this to really work, I feel like I'd have to sit there going: "Okay, same thing but bring the lower mids up a couple of dB. Try it with a little more air in the top end. Compress those drums, but try not smear the transients on the snares. No wait, go back. Try smearing those transients a bit more."

At that point, I'm not sure "mastering as prompt engineering" is worth it. Though I totally agree that things like LANDR and Ozone are great for quick-and-dirty work.

I've been using Ozone over the last few years for my home masters and even with all the AI and bells and whistles, it's nowhere near as good as a master from a studio who uses analog gear. I will also admit my ears are just not there yet. What it is good for is cutting a demo to send out to labels or DJs to play at shows.

We need a Blender in the DAW space. Node-based with python scripting. That could be a powerful tool.

LANDR was released in 2014; AI has been mastering music in this way for a decade..

Mastering Assistant in Logic is surprisingly good, but I do still find that I have to tweak the chain going into it to get the best results out. It's remarkably light on the compression so adding a little bit, along with some EQ, before it goes into MA definitely helps on the kinds of music I make.

This will be baked in to apps like Logic Pro or Abelton very soon

The lyrics of Hair Metal Renaissance promised me a guitar solo, but reneged. I want a pony!

Since music is mathematical I assume that AI will be much better at creating music in the long run than humans will be.

I don't know about that. Music is as much mathematics as playing basketball is physics. Musicians I have met, just do it completely intuitively, they can just feel it. Yes you can describe all notes, their relationships and their resonances with mathematics, but you can do that with anything and everything. That is the very point of mathematics - to describe things.

I think it is odd to say that something is mathematical, when all things can looked at through the lenses of mathematics.

Also for tasks that have clear indications of varying degrees of success. Like throwing a ball in a basket. Yes absolutely computers and AI will in time do them better than humans ever will.

But for music once you make a song that is perceived as "good" or "not bad", there is no such thing as better or worse, it becomes entirely subjective. So I do not know if it is possible for anything to be better than something else. For composers and music makers we often assign celebrity status and perceive some greater than others, but really the music some create is not better than the rest.

Maybe AI will be able to demonstrate technical prowess beyond human ability like be able to instantly write down all 15 sounding notes in a given beat in a song. But it does not make sense for it to better at creating music than humans in general.

music is one of the most subjective things in life, and what one person considers good or meaningful music is based entirely on non-mathematical forces like culture, early child exposure to music, and personal taste. Music only is mathematical in the same way everything can be represented by mathematical concepts. While I have no doubt ai will be able to make presentable music in the near future, music is about expressing the human experience, which is something ai can't do. Treating music like its a product that can be perfected shows a deep misunderstanding as to why people find value in music

I really like hacker news but the users here have very strange perspectives on art.

There are elements of music that are mathematical, but not all of it is.

This is why we don't listen to MIDI tracks all day they are a purely mathematical interpretation of a song and they're mostly terrible.

If music is mathematical, then what isn't mathematical?

It is useful to keep in mind the difference of 'music' and 'musical industry'. Musical industry relies on 'celebrity' and social interactions. It relies very little on its efficiency or performance at 'making music'. People want to hear what person 'X' is singing, and what 'story' they are telling.

To the very few that care only about the sound waves, AI will triumph because it will craft music that is perfect for that specific persons taste, and even include other parameters like tone, mood, state of mind, place of residence, weather...

We have computers and computer science because we can represent things mathematically. This is not insightful nor does it mean some software can automagially engooden things.

I was going to respond that much of music is also poetical, but AI is actually decent in that realm too - albeit it with a lot of supervision.

From 2019: https://www.youtube.com/watch?v=EPs6wdM7S3U

There is a thing in dance music where some DJ or producer will find an interesting new sound or style, and then other DJs and producers will produce sound-alike knock offs, and that will be interesting for _a while_, but eventually people will get bored of it and nobody will buy that kind of music again.

An AI might sort of make producing knock-offs of popular styles easier, but it'll also probably hasten the decline in popularity of those styles as well. One popular song is great, 3 or 4 copies of that style might still be good, but 10,000 indistinguishable copies of it -- meh.

I think what you might actually see is something like this:

A dj comes up with a new sound, and people might like ask an ai to produce a whole mix of songs similar to that record, just for them, that nobody else will ever listen to. Or even better -- a DJ might come up with a new sound and _sell an AI themselves_ that can produce songs like that on demand.