AI can now master your music | Hacker News Reader

AI can now master your music

arstechnica.com

48 pointsdudewhocodes50 comments

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!

> untrained ears will still not make "professional-grade" masterings

There's quite a few people who hate the modern mastering, so... maybe that's not a bad thing? Also considering the pro grade masters are often aimed for wildly different setup than the cheap earbuds many people will use. I'm not trying to say pros don't know what they're doing. But also, I'd love to see people play a bit more. (Still dreaming of a real release format that contains the raw elements + effects stack)

Then again, I'm a weirdo who honestly prefers some accidentally preserved band practice recordings with noise and mistakes and raw energy to their released official albums.

the idea is that if it sounds good on the 'wildly different setup' (i.e. posh speakers in controlled environment) then it will also sound good on cheap earbuds... but not the other way round.

to me the whole 'mastering AI' stuff misses at least some of the point of what mastering really is, but i suppose it'll be useful for people doing stuff on their own who wouldn't pay a proper mastering engineer anyway.

Sometimes true... Sometimes you get Christopher Nolan, with his own quote:

> Because you can make a film that looks like anything, you can shoot on your iPhone, no one’s going to complain. But if you mix the sound a certain way, or if you use certain sub-frequencies, people get up in arms.

well yeah, "good" meaning "as the artist intended"... that's the other thing with this AI stuff, there is no objective definition of "good sounding record" unless you just want everything to sound like Steely Dan or whatever...

I don't think it's weird at all. Art and culture generally are exercises in communication. Sometimes fine craftsmanship or technical prowess and spectacle is an important part of the message. Sometimes it simply is not. People get excited when their favorite sports team tries a daring play, whether it's on an 8k monster display or their grandma's black and white set from the 70s. More pixels won't make you cry any harder in Titanic.

Kind of reinforces the idea AI is going to lead to mediocrity in a lot of industries. I feel like our food, like the stuff at the grocery store is a great example. Our bread is made on a production line. It's good but not great. I feel like the real bummer is the option to get great bread tends to disappear overtime as everyone moves to the 80% solution and the training and economics needed for the artisan solution disappear.

I predict the cost to make an 80 percent solution will plummet, which is great for consumers.

Essentially, consumers now might pay X amount to get only a 40% solution (probably unusable and not something a company would even offer), but with AI, the same X gets you an 80% solution.

The price of >80% solution will remain the same, until AI improves.

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.

I think it's probably more if you are not aware of tuning stereo width or stuff like that, it'll get you to a nice "default". But obviously if you're more knowledgable you might tune this yourself.

For loudness there are certain "standards": https://support.spotify.com/us/artists/article/loudness-norm... - it just makes distribution of your music much easier.

There are probably (or soon will be) different presets for mastering a rock track or an EDM track or a jazz track, with different targets for dynamics, loudness etc.

A music genre is a "standard"

also messing about with stereo width is not (in my experience) a 'default' move in professional mastering. it would be done as a corrective, but only with the say-so of the artist.

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.

That's more or less Reaper though

Not open source and lacks a ton of features like editing sound waves

For node based workflows, check out Max or Pure Data.

https://cycling74.com/products/max

https://puredata.info/

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.

It's been a while since I've made music, but my secret sauce has always been Wave's Abbey Road TG mastering chain (not "AI" but it has tons of nice presets for different genres). I need to test this Logic Mastering Assistant out though to see if they play nice together.

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

Mastering Assistant in Logic, as mentioned in the article:

https://support.apple.com/guide/logicpro/mastering-assistant...

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

Also, AI may now be able to master your song, but if you don't make the guitar loud enough in the mixing stage, it will still sound strange...

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

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

Nothing i guess. Ask anyone who knows music theory, there is a ton of math in music.

Source: I randomly socialize with a professional classical musician that does both playing and composing/analysis.

Well. What I mean is that people have used mathematics in order to create music for a long time. But you're correct that everything is mathematical in a sense.

People tend to like the same pieces of music, so there is obviously specific patterns that people generally like. AI is good at finding patterns so in the end AI will become good at creating music.

I think.

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...

That sounds quite bleak, we're all really going to be living in our own worlds soon aren't we? "Hey did you hear the new album perfectly tailored for me yesterday?...of course you didn't".

The music, I guess. But there are people who like the artist - their interviews, shows, you name it. Not sure how's that going to map on AI-generated music. And no, avatars don't cut it - they are known to be fake so whoever appreciates Dolly Parton for what she's doing will not feel for some avatar issuing the correct words.

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.

The instruments are also mathematical, frequency ratios, spectrums, envelopes ...

MIDI sounds bad because while the notes being played respect a mathematical structure, the MIDI instruments spectrums do not integrate properly into said structure.

I'd really like to understand what you mean by this?

Yes of course, mathematics is used in the design of the instrument but it's not really a "mathematical" / precise device in the same way a calculator is, nor does a good musician conform to mathematics only when expressing themselves using the instrument.

In traditional musics, people use stones which when hit will make a sound that might conform to some frequency, but I wouldn't think the stone is "mathematical"?

I completely understand math is involved in Western and Eastern musical theory, but a saxophone being "mathematical" ? My trumpets and saxophones will only play the correct tones if I work with it to do so.

Disclaimer: been playing sax, trumpet since I was 6.

“My trumpets and saxophones will only play the correct tones if I work with it to do so.”

I have no opinion either way, but this conflates mathematical with automated, doesn’t it?

what is automated about having to change the shape of your mouth and strength of your exhalation to produce the right tone ?

There is a mathematical structure in the harmonics of a piano note. There is a different one in the harmonics of a sax note. Yes, it's not 100% structure, there is noise too.

This is why the "timbre" of a piano is different from the timbre of a sax, and why the same C note might fit in a song if it's played by a piano, but not in the same song if it's played by a sax.

If you take a higher level view, the mathematical structure of the notes (the chords, ...) needs to fit the mathematical structure of the harmonics of the instruments (the timbres) to get a pleasing sound. Put another way, the timbres of some pairings of instruments might clash on a particular sequence of notes, so you need to search for a different instrument pairing which is harmonious on that sequence.

And then you have synthesizers, which are very mathematical in the way the timbre is generated (oscillators, filters, ...), but this allows you to exactly fine tune the timbre so it "fits"

It seems a bit philosophical to me though, what came first, the music or the maths?

We can use mathematics to describe the theory behind music but the theory isn't the music itself, nor are the instruments, "mathematical".

When a great performer plays live, they're not generally perfectly reproducing what's written in the musical notation either.

Probably not a straightforward thing to argue about, as I said, I think it's mostly a matter of viewpoints.

There was a paper which studied the neurological basis of music.

One of the function of the brain is to predict the future. Neurons and brain systems are rewarded for correct predictions and punished for incorrect ones.

The main theory of the paper was that music is pleasant because it's repetitive nature and simple patterns makes it easy for the brain to predict, thus there is a lot of reward to extract from it. They focused on modern trance music and ethnic african drumming, both which feature very long very repetitive sequences, which reward the brain so much that they can induce a trance like state.

Not any random combination of physical objects is a musical instrument (yes, avantgarde music exists, ...) The sax is built the way it is because it generates a sound spectrum with harmonic relations between the frequencies (integer ratios, ...). Those who originally built the sax didn't use math to design it, but it so happens that the result respects certain mathematical ratios, because those ratios are highly predictable and pleasant to the brain.

A funny example of this I came across recently.

There's a video game called Sonic Chronicles--just before it was completed, the developers were prevented from using the soundtrack they had made and were forced to recompose new songs for the entire game for reasons that remain murky. The result are some almost incomprehensible pieces of minimally instrumentated MIDI music like [this](https://www.youtube.com/watch?v=6o47N-aYd08).

Yet despite sounding like complete ass, those songs were actually lifted from previous Sonic games, where they sound [completely different](https://www.youtube.com/watch?v=t4-DQ4hGRJo) despite being structurally identical.

Well, not all MIDI sounds bad. Have a listen to the Addictive Drummer demos for example. It can sound pretty damn good / real. Still, your point is well taken. MIDI guitar emulations for example are still pretty dang bad. One wonders if this is because a guitar is a more continuous vs discrete instrument compared to drums. Even with all of MIDI’s rich parameters, it’s pretty tough to describe pick angle, materials, etc.

Also, I wonder if there will be a natural “backlash” against generic AI music and towards stuff that has more human imperfection. But then again even the AI will learn to imitate the imperfections so what exactly will make a piece more human in the long run?

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

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.

Question: could you enjoy a song if you knew absolutely nothing about the artist? Let's make it instrumental so you can't judge the voice.

From what you say I infer you couldn't, you need to know about their life, their struggles, their views on things, etc.

Some people can actually enjoy just the pure sound. Electronic club music tends to be like that, where if you listen it in a club or on the radio in a long mix you might have no idea what song is being played, the vocals and lyrics are ultra-minimal so no "human expression" there.

of course I think you can enjoy music as just "pure sound". I have no doubt that generative music models will make perfectly adequate music, and will almost certainly have the capacity to make songs that are great and moving. But just like visual mediums, music is more than just a file. Music is cultural. Electronic music is absolutely human expression, even if it isn't making some grand statement on society or the human condition it exists in context of other electronic music tracks. By simply having someone feel the drive to create a song and making creative decisions about how that song should sound and feels it makes a comment on the genre of electronic music, and contributes to the cultural project of music.

thanks for the thoughtful comment and I hope you have a nice day.

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.

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.

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.

Reply on news.ycombinator.com