Just take a few pop songs written by professionals then do some descent with modification - and dump them all onto youtube. Cull the losers - rinse and repeat.
We may also get machine learning systems which take in recordings of existing singers and generate the data needed for a singing synthesizer. Automatic cover bands, at last. It will be worth it just to really annoy the music industry.
Given the relatively short, finite length of the average pop song; given the well known, restricted inputs; given that there is a general formula for successful pop songs; not only can AI write pop songs, it can (and will) write every possible pop song. At this point it's merely a question of a short amount of time, sub ten years perhaps, twenty tops.
The creation part will ultimately end up being the trivial aspect. Figuring out which of the billion songs AI create are any good will be the more difficult, time consuming filtering / discovery aspect (that will process partially at human-speed).
AI will create vast amounts of content/goods in the relatively near future (of all types, from VR worlds to pop songs to physical things automatically made by printing). A large industry will spring up around filtering that, deciding what's good, to then serve it up to people for popular mass consumption.
This is the most insightful comment. Pop music tends to be particularly susceptible to the impression that it can be copied, because it seems to fit the idea more than any other genre (that is debatable of course) that successful music is a template. But even in pop, things progress. AI has been shown to be pretty decent at copying and modeling. Creativity? Not so much. We've been making the equivalent of HMM-based "music in the style of (i.e. generative stochastic processes modeled on) Beethoven" since the 60s, but such a design is to going to write the next hit. Or if it does, it won't write the following 10 by itself. You would think it's not true from how the radio presents it to you, but the fact is that people (eventually) get bored of hearing the same thing over and over.
> You would think it's not true from how the radio presents it to you, but the fact is that people (eventually) get bored of hearing the same thing over and over.
Evidence is against you.
The number of radio stations that play something "new" are VASTLY outnumbered by the radio stations that play the same things over and over and over.
If we start removing things which sound almost identical (See: modern "bro" country, etc.), the number of places where you can get exposed to new music dwindles to effective zero.
> The number of radio stations that play something "new" are VASTLY outnumbered by the radio stations that play the same things over and over and over.
That's true. I'm not really convinced it's a reflection of what people want though, so much as a reflection of the bare minimum necessary to satisfy listeners.
I feel it's a bit of a self-fulfilling prophecy though --- it sells because it's there.. but everyone complains about it.
And it's there because it's easier for companies to repeat the same thing than to make something new. It's not necessarily a reflection of what people want. It's more like.. the lowest common denominator.
I think in the end, play time and music popularity are much more complicated social phenomena than whether people think a song "is good" or not.
However, maybe you're right in that this phenomenon could play to the benefit of AI-generated music. But I really think it will hit the limits of people's patience, quickly.
>We've been making the equivalent of HMM-based "music in the style of (i.e. generative stochastic processes modeled on) Beethoven" since the 60s, but such a design is to going to write the next hit. Or if it does, it won't write the following 10 by itself.
This is important to recognize, yes, but it's also important to realize that for every 1 AI agent that writes a hit, there's going to millions others (or more) crunching away at songs that aren't hits (like musicians currently, but many more, working more hours, and pumping out many more songs each).
The next 10 hits don't need to all come from the same AI musician to "take over" pop. They just need to come from 10 of millions others all competing with each other (and humans) for those top spots.
Even with current technology, this really isn't so much a music generation problem as it is a filter problem asking "how do you know which of these billion (generated or not) songs are going to be hits?".
I hope you are being sarcastic :) Even though pop songs aren't that imaginative, exponential growth means causes problems. We can't count all ways of ordering 52 cards using all the computers in the world in the history of the universe, and I bet one could come up with an algorithm for creating pop songs, one per ordering of a deck of cards.
Nope, not being sarcastic. You're radically overestimating the range of potential pop songs (much less the range of possible good songs, which is far, far, far smaller).
Human ears & taste/preferences drastically, inherently reduce the set. It's a deck of 9 or 10 cards, not 52.
When I look for new music in a particular style (let's say trap), I can discard 90% of new songs with just 2 sec snippets. Very few songs get 30 sec of listening, and even fewer a full listen. I would say I'm pretty good at it (the ones I like tend to become hits).
If I can filter 90% in 2 sec, you can probably machine learn it. Of course, if after you filter out the easy stuff, you still have heaps of songs, there will be a problem.
But I think the solution will still be some form of machine learning. So we'll have a long automated pipeline filtering out stuff, and only surface for human rating a minuscule amount. Gems will be thrown away, but it won't matter.
Me, I wait for the day when algorithms will invent a new genre (house/trance/drum&bass/dubstep/trap/???).
And one thing that keeps coming to my mind. Can "good music" be absolutely rated? Will eventually advanced computers consider our music good? Will they "like" completely different music, which sounds like garbage to us? Or to put another way, how much of music depends on general intelligence, and how much on the specific realization of said intelligence (brain/culture)
Really? Yes there are some intervals that we have discovered/learned in our audio spectrum, but the fact that you can so confidently discredit something highlights your ignorance. Unfortunately, life ain't no engineering problemo
Having listened to the song linked in the OP I was blown away, but then:
> I ought to note that a human being is involved in all this: the French composer Benoit Carré, who took the raw data generated by the CSL and matched it to audio from existing recordings [..] Benoit then arranged, produced, and mixed the tracks that resulted from this collaboration. Carré has also written the words.
I feel like there's a big difference between "AI creating music" and "AI as an assisstive tool for creating music".
But in the early 90's I sat in on a seminar by a music prof at my university, he was working on AI generated music. His vision was that it would be an assistive device; basically labor saving for composers.
At the time his students were doing automated compositions and I was blown away by the quality at the time. Definitely his students were already at the stage of partially composed music - basically just a songwriter who needed a good arranger / producer.
Given some of the work I've read on RNN's, I think this is really close to being fully automated. 5 years, maybe, and stuff like this will be fully automated.
I think this is the future of AI in music, not just pop.
It's trivially easy to write a mediocre song. All it takes to turn a mediocre song into an above-average one is a clever turn of musical phrase; be it melody, harmony or rhythm or all three (typically it's all three). As a songwriter you're constantly generating options for you to choose from. "What chord goes well after this one but doesn't sound too predictable? What direction can I take this melodic idea in? My chorus is boring; is there something I could do rhythmically that would make it more interesting?" You're constrained by your own imagination, which itself is constrained by what kind of mood you're in, what you've eaten or drunk, energy levels, etc.
How nice would it be to have a super-charged co-writer that could spit out 20 alternatives that you would never have thought of. They don't even need to be perfect alternatives; simply enough to jog your head into a different frame where you can proceed to refine and build from.
Yeah, excepting that a label exec would immediately scoop them up to be produced as remix singles, to be released in stages over the next six months on iTunes, Amazon, and Google Play for $1.29 apiece, whether you liked them or not.
Another point - back in the 80's, when New Age was getting popular, there was a New Age radio station that featured a format without radio personalities - it was really to my knowledge the first totally programmed station.
One music instructor got together with a couple friends and did assembly line New Age - they recorded a bunch of A sections, a bunch of B sections, and pasted them together A-B-A, A-A-B-A, that kind of thing, and turned around and got radio play on these compositions.
My point is that this kind of thing has been going on for a while. Now we're learning how to make computers do it.
That's interesting, but in that scenario, how important is the composer really?
Couldn't we just connect the variable output of the AI to some kind of Mechanical Turk-style focus group, feed those results as inputs back into the AI, and repeat that process millions of times? Might that produce a better song than an individual composer?
I've seen interviews with some modern artists, and they'll often take inspiration from decades-old hits (which are essentially de-risked melodies) and apply them to a modern template. I'm assuming the bigger labels already focus-group the output and make revisions. This seems like something that could eventually be automated.
sounds like the ol' stone soup trick, where its all the extra ingredients which make the magic work, not the AI. I'd reckon a talented producer could make a good song out of pseudo-random notes. Unless the original untouched AI track is released, I have to be skeptical
Pop evolutes. With machine learning we can create, say, every single combination of reggaeton, but every decade new styles appear. AI leveraging real creation requires AGI.
Some years ago I was backstage at a major rock event the day before the performance. One of the management types was explaining to me that they had two sets of everything, and three crews. One crew was at the current venue, one was at the previous venue, tearing things down, and one was at the next venue, setting up. I asked "so why not have two sets of performers and double your revenue? 'Cats' has two road companies. Barnum and Bailey circuses have two road companies. There are bands that replaced all their members over time and kept the name. It's all about the branding." A more senior manager, overhearing this, looked very thoughtful.
It's been done in Japan. AKB48 has three teams. They're organized like a sports league, with farm teams and regular turnover.
I'm not all that familiar with them, but isn't Gorillaz an entirely fake band? I mean, they don't have real human members, it's all virtual musicians.
In a sense, all the boy bands are basically the same thing. One runs its course, another springs up with the exact same formula. It'd be much more efficient without having to deal with human performers.
I think this franchise music thing is going to happen.
I have essentially zero experience with AI, but a solution to writing songs with neural networks to me seems like it would require the following:
- Parse the different instruments from raw audio
- Parse the timing pattern from each audio layer
- Turn each layer into a roughly similar midi sequence
- Develop a pattern of relationships between the different types of layers and the patterns that are commonly found together (ie. a closed hihat on 1/8ths or 1/16ths is often accompanied by a sub bass pattern of some kind in trap music)
- Using the knowledge of these sorts of relationships, build up a song by specifying some parameters ("I want a hip hop beat with sub bass, snare, and hi hats")
A lot of pop songs go through the hands of several different producers, each of whom might work on a specific element for a desired effect. In many cases these producers are selected for specific sounds deemed to be chartworthy. I would be very surprised if some of them are NOT using AI trained on extensive databases with the latest most popular and most lucrative music.
I recently worked on a project that attempted to find harmonizing chord progressions for pop songs, given a melody. Since pop music has such predictable structure, I think AI generation is well within possibility.
Has anyone fed the Real Book into a Markov Model (or I guess RNN these days) and seen what gets spit out?
All the music examples I've seen have been people feeding tracks because...people tend to have lots of MP3s lying around. I'm curious as to what you get when you work directly with the notation which conveys a lot more structure than a waveform.
50 comments
[ 3.2 ms ] story [ 117 ms ] threadJust take a few pop songs written by professionals then do some descent with modification - and dump them all onto youtube. Cull the losers - rinse and repeat.
We may also get machine learning systems which take in recordings of existing singers and generate the data needed for a singing synthesizer. Automatic cover bands, at last. It will be worth it just to really annoy the music industry.
I could have sweared it was natural. Especially the Spanish sample.
The creation part will ultimately end up being the trivial aspect. Figuring out which of the billion songs AI create are any good will be the more difficult, time consuming filtering / discovery aspect (that will process partially at human-speed).
AI will create vast amounts of content/goods in the relatively near future (of all types, from VR worlds to pop songs to physical things automatically made by printing). A large industry will spring up around filtering that, deciding what's good, to then serve it up to people for popular mass consumption.
Imagine a world where the Police never broke up?
Evidence is against you.
The number of radio stations that play something "new" are VASTLY outnumbered by the radio stations that play the same things over and over and over.
If we start removing things which sound almost identical (See: modern "bro" country, etc.), the number of places where you can get exposed to new music dwindles to effective zero.
That's true. I'm not really convinced it's a reflection of what people want though, so much as a reflection of the bare minimum necessary to satisfy listeners.
I feel it's a bit of a self-fulfilling prophecy though --- it sells because it's there.. but everyone complains about it.
And it's there because it's easier for companies to repeat the same thing than to make something new. It's not necessarily a reflection of what people want. It's more like.. the lowest common denominator.
I think in the end, play time and music popularity are much more complicated social phenomena than whether people think a song "is good" or not.
However, maybe you're right in that this phenomenon could play to the benefit of AI-generated music. But I really think it will hit the limits of people's patience, quickly.
This is important to recognize, yes, but it's also important to realize that for every 1 AI agent that writes a hit, there's going to millions others (or more) crunching away at songs that aren't hits (like musicians currently, but many more, working more hours, and pumping out many more songs each).
The next 10 hits don't need to all come from the same AI musician to "take over" pop. They just need to come from 10 of millions others all competing with each other (and humans) for those top spots.
Even with current technology, this really isn't so much a music generation problem as it is a filter problem asking "how do you know which of these billion (generated or not) songs are going to be hits?".
I hope you are being sarcastic :) Even though pop songs aren't that imaginative, exponential growth means causes problems. We can't count all ways of ordering 52 cards using all the computers in the world in the history of the universe, and I bet one could come up with an algorithm for creating pop songs, one per ordering of a deck of cards.
Human ears & taste/preferences drastically, inherently reduce the set. It's a deck of 9 or 10 cards, not 52.
It's 52!, or 80658175170943878571660636856403766975289505440883277824000000000000.
Enumerating them is where it starts to get tricky.
What would you say that is?
Your answer has to be detailed enough that a computer can understand it.
There's a lot of patterns to pop songs but there's patterns in all music. I haven't seen any function that inputs x and outputs a pop hit.
If I can filter 90% in 2 sec, you can probably machine learn it. Of course, if after you filter out the easy stuff, you still have heaps of songs, there will be a problem.
But I think the solution will still be some form of machine learning. So we'll have a long automated pipeline filtering out stuff, and only surface for human rating a minuscule amount. Gems will be thrown away, but it won't matter.
Me, I wait for the day when algorithms will invent a new genre (house/trance/drum&bass/dubstep/trap/???).
And one thing that keeps coming to my mind. Can "good music" be absolutely rated? Will eventually advanced computers consider our music good? Will they "like" completely different music, which sounds like garbage to us? Or to put another way, how much of music depends on general intelligence, and how much on the specific realization of said intelligence (brain/culture)
That apparatus already exists and is quite functional. Are you fully in charge of choosing what music you discover and end up liking? Are you sure?
Really? Yes there are some intervals that we have discovered/learned in our audio spectrum, but the fact that you can so confidently discredit something highlights your ignorance. Unfortunately, life ain't no engineering problemo
> I ought to note that a human being is involved in all this: the French composer Benoit Carré, who took the raw data generated by the CSL and matched it to audio from existing recordings [..] Benoit then arranged, produced, and mixed the tracks that resulted from this collaboration. Carré has also written the words.
I feel like there's a big difference between "AI creating music" and "AI as an assisstive tool for creating music".
At the time his students were doing automated compositions and I was blown away by the quality at the time. Definitely his students were already at the stage of partially composed music - basically just a songwriter who needed a good arranger / producer.
Given some of the work I've read on RNN's, I think this is really close to being fully automated. 5 years, maybe, and stuff like this will be fully automated.
It's trivially easy to write a mediocre song. All it takes to turn a mediocre song into an above-average one is a clever turn of musical phrase; be it melody, harmony or rhythm or all three (typically it's all three). As a songwriter you're constantly generating options for you to choose from. "What chord goes well after this one but doesn't sound too predictable? What direction can I take this melodic idea in? My chorus is boring; is there something I could do rhythmically that would make it more interesting?" You're constrained by your own imagination, which itself is constrained by what kind of mood you're in, what you've eaten or drunk, energy levels, etc.
How nice would it be to have a super-charged co-writer that could spit out 20 alternatives that you would never have thought of. They don't even need to be perfect alternatives; simply enough to jog your head into a different frame where you can proceed to refine and build from.
One music instructor got together with a couple friends and did assembly line New Age - they recorded a bunch of A sections, a bunch of B sections, and pasted them together A-B-A, A-A-B-A, that kind of thing, and turned around and got radio play on these compositions.
My point is that this kind of thing has been going on for a while. Now we're learning how to make computers do it.
Couldn't we just connect the variable output of the AI to some kind of Mechanical Turk-style focus group, feed those results as inputs back into the AI, and repeat that process millions of times? Might that produce a better song than an individual composer?
I've seen interviews with some modern artists, and they'll often take inspiration from decades-old hits (which are essentially de-risked melodies) and apply them to a modern template. I'm assuming the bigger labels already focus-group the output and make revisions. This seems like something that could eventually be automated.
If you think about it, musicians work at low level - assembler or even raw mechanical inputs. They could use a toolchain.
[1] http://songsmith.ms/
Some years ago I was backstage at a major rock event the day before the performance. One of the management types was explaining to me that they had two sets of everything, and three crews. One crew was at the current venue, one was at the previous venue, tearing things down, and one was at the next venue, setting up. I asked "so why not have two sets of performers and double your revenue? 'Cats' has two road companies. Barnum and Bailey circuses have two road companies. There are bands that replaced all their members over time and kept the name. It's all about the branding." A more senior manager, overhearing this, looked very thoughtful.
It's been done in Japan. AKB48 has three teams. They're organized like a sports league, with farm teams and regular turnover.
Had; Barnum and Bailey folded in May.
In a sense, all the boy bands are basically the same thing. One runs its course, another springs up with the exact same formula. It'd be much more efficient without having to deal with human performers.
I think this franchise music thing is going to happen.
- Parse the different instruments from raw audio - Parse the timing pattern from each audio layer - Turn each layer into a roughly similar midi sequence - Develop a pattern of relationships between the different types of layers and the patterns that are commonly found together (ie. a closed hihat on 1/8ths or 1/16ths is often accompanied by a sub bass pattern of some kind in trap music) - Using the knowledge of these sorts of relationships, build up a song by specifying some parameters ("I want a hip hop beat with sub bass, snare, and hi hats")
Doesn't matter. There is no value in the songs, they are easy to write and many are written in a day with the premise of becoming the next pop song.
Pop is about distribution and marketing. Control it and your songs will be the next thing played everywhere.
Seems like this is the broader insight to the whole "can AI make art/music/whatever" debate that many are missing.
https://luckytoilet.wordpress.com/2017/04/25/ai-project-harm...
All the music examples I've seen have been people feeding tracks because...people tend to have lots of MP3s lying around. I'm curious as to what you get when you work directly with the notation which conveys a lot more structure than a waveform.