Musical Analysis at Moogfest

Stuff they Don't Want You to Know. Noel went to mog Fest in and and got the chance to talk to a whole bunch of really cool people, including Alexander Lurch and we'll hear more about that a little bit later in this podcast. Mog Fest ostensibly is about music and technology, but it actually involves lot lots of other stuff to not just not just those two already broad fields, but other ones as well, including elements of philosophy and

and even particle physics. Will have an episode in the near future that will include some elements from uh interviews we had with folks from the Large Hadron Collider. So mog Fest has all sorts of really smart, talented people getting together and having these incredible symposia and and and performances, And so Noel was able to go and talk with someone about some really cool stuff, and that kind of

ties into what I wanted to chat about today. You know, once upon a time here at How Stuff Works, we had a show called Stuff from the B Side, and this was a podcast all about music. Episodes focused on everything musical, including elements that are more general concepts or philosophical ideas. And music and technology are two things that

really do closely tied together. After all, almost every musical instrument is some form of technology, ranging from the relatively primitive versions of certain percussive instruments all the way up to high tech digital rigs. So I thought it might be cool to revisit music and tech and look at

a particular subset of it, musical analysis and music generation. Now, music analysis and technology are also related in that we now have various automated recommendation engines that will suggest music for us to listen to based upon what we've already said we enjoy. Now these engines look for new pieces of music that in some way match criteria we seem

to find appealing. We have indicated to that service that we like that particular type of music, so it starts to try and find matches that kind of follow in the same lines. As they become more adept at figuring out what qualities we really enjoy, they can hone in on songs that appeal to us, perhaps even changing them up based upon other criterias, which is the time of

day or an activity. We're doing so, for example, with Google Music, and this show is not sponsored by Google Music or anything of that nature, but it will detect if I'm on my way somewhere. It might suggest music that would be conducive to a trip, or if it knows that I'm at the gym, it may suggest music that's good for keeping my heart rate up, stuff like that. So we'll just imagine a hypothetical situation. I've just woken up and the recommendation engine might find some peppy music

to get me on my way. So Google Music is saying, hey, it's Monday morning, you need all the help you can get. Here's a radio station based off the song Walking on Sunshine by Katrina and the Waves. And then my phone detects that I'm going to the gem, so then the music engine switches to the song's meant to keep me moving at a particular pace while I desperately try to find the exit to the gym. I'm sorry, I'm uh

to actually work out. So in that case, it's probably you know, something with a nice driving beat a good tempo to it. These are basic things that music engines can do now, but the reason they can do them at all is because of music analysis. This isn't always done in an automated fashion. In fact, automating music analysis is pretty tricky. Sometimes it relies instead on just a lot of work, and that's work done by real, live, human beings. So let's take the Music Genome Project for example.

This is the database that the internet radio service Pandora relies upon when it creates a radio station based off an artist or a song that you've submitted as the seed for a new channel. For more than ten years, Pandora's staff have analyzed and categorized music, breaking down songs into all the basic components, which they call genes. These

are the elements that make songs what they are. And I find this approach both fascinating and and a little odd, because in a way, it seems a little weird to take a really awesome song. Let's say it's um Blue Oyster Cults, Don't Fear the Reaper, one of the best songs ever written. And then you have to sift it down to all those little basic components, those genes that make up that song. It also reinforces this notion that a song is more than just the sum of all

its parts. If you were to look at those components and attempt to make a song that included all of them, I bet it wouldn't be half as awesome as Don't

Fear the Reaper. So you take a song, you identify all these different qualities of it, and may involve things like the tempo of the song, the the the structure of it, as far as versus and choruses are concerned, the whether what kind of vocalists there are, what kind of instruments are used, all of these different individual, tiny components of the song, and you put them into say spreadsheet, and that represents the collection of genes that are possessed

by Don't Fear the Reaper. You take that same collection, you give them to a musician and say, I want you to write me a song that has all of these components in it. Well, again, probably not gonna get Don't Fear the Reaper. You'll get something, and maybe it will be good. Maybe it'll even be better than Don't Fear the Reaper. I doubt it, but yeah, there's there's something magical or apparently magical about music that transcends the

quantitative elements that we can list now. Pandora's Music Genome project identifies four hundred fifty different musical attributes or genes. They include lots of different types of data. Some of them are relatively straightforward, such as does the song have a vocalist? If it does have a vocalist, is it a male vocalist or a female vocalist? Are there multiple vocalists?

Then starts getting way more granular. So if a song has electric guitar, for example, there might be a subset of information about that, such as how much distortion is on that guitar? Does it have a lot of distortion in this song or not a lot? And so you start to subdivide down the line. Same thing is true for other instruments as well. Now, not all songs have the same number of genes, meaning some genres of music are actually easier to describe with a fewer terms than others.

For example, rock songs have about one fifty genes. You can break down your rock song into about a hundred fifty different little individual components. Rap songs are more like three d fifty. So that indicates that there are gradations and variations between different songs within the same genre. Uh So, to make a recommendation engine, you first have to put all the music within the library. Through this process, you need to identify the important qualities that make the music

what it is is. And you could use something like a spreadsheet and you lay it all out, and then when someone wants to make a new radio station off of a song, you can use that song's genome all the jenes listed for that specific song to guide a decision engine to pick other songs that are similar to the first one within a certain degree. So you could

set this dynamically in your search engine. Right Like, let's say that you are the one designing the new, latest and greatest version of Pandora, and you've got this enormous

database of music that's all been analyzed by professionals. We're talking about actual musicians and musicologists who have listened to the music, broken it down into its basic elements identified all of them, and someone has joined your service and they say, I'm going to make a radio station based off the song, Uh, the statue got me high by

they might be giants. You would end up accessing the database, pulling the record for the statue that got me high, looking at all the genes that are associated with that, and then you would look for a certain percentage of similarity with other songs, like are there other songs that have the same genes as this song does? If so, serve it up see if the person likes it. You might set the threshold higher or lower. If it's a

song that's particularly avant garde. There may not be a lot of other songs that strongly resemble your original, so you have to kind of play fast and loose with this. Now, an important component of this service is user feedback. Services like Pandora nearly always include a method for users to indicate if they like or don't like a particular song. The recommendation engine uses that data to fine tune its selections.

No two songs are going to be exactly alike, so it may be that the ways the news song deviated from your seed songs format were the parts that made you detest it, So it could have been that the the the engine said, well, this song resembles the seed song, the original tune of the way. Let's serve it up and you listen to it for like three seconds, you say, no, this is this is not what I want. You give

it a thumbs down. The algorithm might say, all right, well, I'm gonna keep note of where it was the same and where it was different from that original song. Meanwhile, I'll serve up this next song that has similarity. And if you say, yeah, that's a good song. I really like it, and you give it the thumbs up, then the recommendation engine starts looking at the differences between the song you said no two and the song you said yeah too, and it starts to identify stuff that you

might not even be aware you don't like. It might be certain elements of songs, and the recommendation engine has figured it out. Maybe it's figured out, oh uh, Jonathan really doesn't like it when there's a clarinet in the song for no reason, but he isn't able to vocalize that he doesn't he's not aware of it consciously, but every time it's popping up he's saying no to that song, So we're gonna We're gonna put the kai bosh on the clarinet from here on out. That was just a

random example. I don't I don't have a hatred of the clarinet, but it is a way for the engine to work with the user in order to get a better understanding of the type of songs that it should serve up to you. Now, there are plenty of other ways to analyze and describe music besides this genetic approach. There are entire courses dedicated to this. Musicology is a rich and interesting field, and some of these approaches go

beyond the components that are directly perceptible. These analytic methods try to capture the essence of the feel of music. For example, if you take a bunch of components individually, you might quantitatively describe the music with accuracy, but you can't capture how they collectively create a particular effect. Perceptual analysis attempts to bring human perception and emotional reaction into account with everything else. But why is the Music Genome

project powered by humans? Why is Pandora using actual human beings to listen to music and then write out all these genes, couldn't you find some easier way? Well? Listening to music and being able to describe its structure beyond some relatively simple angles is a particularly tricky computational problem. It's something that's easy for humans and hard for machines.

In two thousand five, Way Chai of m I T wrote a paper titled Automated Analysis of Musical Structure in which she laid out the challenges of creating an automatic approach to analyzing music. Her pay Earth is nineties six pages long, and that kind of gives you an idea of how complicated a problem this is that we're talking

about here. China's team relied on music cognition, machine learning, and signal processing to segment and analyze pieces of music, with the goal of isolating and analyzing the recurrent structures of a piece. You know, the whole verse, course, verse, all my fellow Pixies fans out there, the chord progression or key changes that are present in music. Identifying parts of a piece that make it representative of the whole. In other words, finding that hook or finding that element

of a song that make it stand out. China's team had to figure out how to make a machine do stuff that we tend to do naturally, even without the benefit of formal musical training. So, for example, I have never taken any class beyond music appreciation, which is about as one oh one as you get, and yet I am able to voke realize certain things about music easily. I can recognize these differences, things that a computer cannot natively do without all and it requires a whole lot

of work. The whole paper is available to read online. It's really interesting. I recommend checking it out. There's a PDF you can just download for free and read over it, and it's fascinating. It delves into not just the programming challenge of creating this analysis software, but also the peculiarities of music itself. For example, what makes one piece of music more memorable than another piece? What element does repetition

play when it comes to making a masterpiece? Was the relationship between music, which, when you get down to it, really is just math and motion and human perception. And I could do an entire episode on Chi's work and what her team developed and how they set out to design this automated system to analyze music, but that's gonna have to wait for a later episode. For now, it's just important to understand the music is something that we're

able to experience in a level that machine just cannot. Now, when we come back from the break, we're going to listen in on an interview that Noel Brown had with Alexander Lurch and learn more about musical analysis and music generation.

But first let's take a quick break to thank our sponsor. Now, Like I said the top of the show, earlier this year, in Producer Extraordinary, Noel Brown took a trip to mog Fest, which was a you know, it's a conference about music and technology and science and lots of other awesome stuff, and he got to speak with a music analysis expert, Alexander Larch And what follows is their conversation. So as a bit of a layman, I interpret a lot of what you do in the field of like generative music.

Is that kind of along the right lines. So um, I would say my book may kind of lead to generative music, but what I'm actually currently focusing on is more analyzing music, so figuring out what's going on in the music. So, um, it might start with you just have an audio signal and you want to know, okay, what is the temple, what is the what is the key, what is the hook line, what is the base doing?

What is the mood of this piece of music? And that is when trying to apply artificial intelligence and signal processing methods to get this information to extract this inflammation

from the signal. So that's something like the hit factories in Sweden would be all about, you know what they're all about, Like it seems that they take a very analytical approach to writing pop songs, where you know, they've got people that are experts in hooks, they have people that are experts in versus, and they have all these kind of human algorithms on like how long everything needs to play for in order to elicit the proper response.

Is it sort of along those lines as well, yes, and so so you you want to find out, um, what kind of makes the songs successful and this might have really many many different factors impacting that. Right. So there's the structure, of course, but there's there's so many other dimensions here that it's really hard to nail it down.

So using using the computer to analyze this, we try to find out more about what's going on and maybe identifying these little things that might make something popular or might give you goose bumps, or something that an example or something that maybe one wouldn't expect might accomplish something like that, or just just like an element that maybe isn't so obvious to the average listener. It's okay, let

me let me think. Like it's it's hard to come up with a very good example that would be surprising to everybody. But it's definitely the combination of tiny things like maybe intonation that is somehow a little bit off, so you would say, or timing is a very obvious thing. If something grooves or not it might have the same rhythm, it might really impact you on a on a completely different level. Right, So these are examples that are maybe

not surprising, but but still um point to the direction. Yeah, is it maybe an element of human human human interaction? Like I think things are too quantized, it's maybe less emotional, whereas when people enter the notes by hand and they're a little bit imperfect, or for example, the singer Adele, there was an article about how she sort of slides into her notes and that gives you goose bumps because it's got this human quality where you sense that raw

human emotion in the same way. Maybe someone who does electronic music makes mistakes and leaves them in and that's what kind of makes it more approachable. Absolutely. I mean, one thing you have to keep in mind is that it's all jover and artists dependent as well, right, so there's there will definitely never be a formula. So if you want to have goose bumps, just do that and

then it looks right, So you can always analyze in retrospect. Okay, this artist has this specific thing thing that he or she does and that makes things so so um fascinating or also that makes you hooked on that, But that might not work for a different genre or for a new song, right, especially because it's also about expectation and

what you already know. So um, I can maybe let a computer compose something in Mozart style, right, and it might be a really good motor piece, but that doesn't mean it really gets you as a listener because you have heard so many Mozart pieces and the original will

still be better. It's it's always an imitation, right, so so then it might actually miss something there, right, Even if the composition itself is very much like Mozart did it, well, so is the end product of your research to make computers better at doing this or are you just interested in kind of you know, breaking down pieces of music and to their based elements. So at the moment, I'm doing exactly that, I'm breaking it down. I I want to be able to let a computer transcribe what's going

on in the music. I want to understand maybe on a perceptional level. So what makes what parameters that you can objectively extract from the audio signal? Um? What impact might they have on the listener? Right? So so how does the listener react to certain um specific characteristics of the music. But this knowledge is then also can most definitely be used to actually generate new music, um, following specific rules that you have extracted from the music and

then create something new. And this is what my colleague Gil Weinberg woks a lot on with his robots that make music. Okay, tell me more about that. And let's not he the mr what it was interested? Right? Yeah? So so there's um he has a robot called him On. So she's a marimba playing robot. Um. So what Also, my my colleague is a lot into jazz, so Simon

plays also a lot of jazz UM. So there's a lot of um interaction on the stage with the live musicians, and the question answer games between what what Simon plays on the marimba and what the musician then plays, and so it's it's constantly analyzed what's being what's being played, and then the robot improvises or tries to um give some answers to that jazz. I mean, you have to listen, you have to be able to follow the leads that you're you know, fellow musicians are putting out there, otherwise

you're not any good exactly. This whole interaction thing is is part of the of the research obviously, and it's not only the music, right, it's only it's also just just it's eye contact and so on. So that's why this robot, even if it doesn't make any sound, has actually ahead where where she can look at specific musicians um and not her head and so on. So you see, you kind of can interact with the robot. So this, this human robot interaction is part of the research as well. Fascinating.

What can you describe the difference between an algorithm that does what you're talking about and analyzes music and one that might create generative music. It seems like there's sort of a crossover between the two, and I'm just I just was probably you could kind of like spell that out a little bit for us. So, UM, in essence, the the algorithm that analyzes music is kind of the information you gain from that algorithm has to feed the

generative algorithm. So, for example, you cannot compose something in classical style if you don't know classical style, right, so you have to learn it from data. That is the analysis part, and then you try to infer models from that. Right.

So you you have all this data, you have you know, um, you have structural data, you have voice leading, you have maybe intonation if it's about performance, and then you try to fix this data into rules, and these rules then would generate music, for example, jazz improvisation or something that. So Brian you know, has has been kind of delving

into generative music lately, and it's actually really interesting. There's a BBC documentary of him kind of showing his methods and he's just using logic and he has these little kind of nodes I guess you could call on the scripts or whatever that can set rules for like a drum part or something like that where it will say, subdivide every other whatever, like any number of things that

you could input like that. Um, I guess are we at a place where that's still just kind of a gimmick or are we Are we really trying to recreate a human mind creating music or is it just kind of a different animal altogether, you know what I mean? Like, I'm wondering, are we really trying to have AI that can compose mozart, or that can place a producer or replace a songwriter, or is it just sort of like

its own thing that's fascinating in and of itself. So I don't think that the goal here is to replace musicians, but I think it's um from a research perspective, Um, giving a machine creativity is a really fascinating topic, right, So is it possible if you just have something algorithm driven, um, that it actually creates something new that it hasn't seen before. Right?

So um I UM, I wouldn't be worried about being replaced, although I mean I could see in the future, like for example, generating elevator music, right, UM that that I can easily see being automatically generated in the future. Um. And there, yes, you would actually the AI would actually replace the human composer and that in that area. But I don't. I don't think that. Um. I think the

the phenomenon of creativity is still not completely understood. Um. And it's with current technologies, it's I think it's really hard to get there. I mean, we do use some random randomizations and so on, so it generates something that you haven't heard before, but well it's random, right, so it's not necessarily an act of creativity here. So so we're trying to get there, but I think it's still a long way to have to create something that is

really creative. It's not getting a Creativity seems to be sort of subjective in and of itself. It's like, does creativity mean that it was created by a human? You know, like is that exclusively what creativity is? And if we have something that is somewhat sentient, can it be creative? You know? I would say that the definition of creativity is mostly subject based. So there's no godlike instance who says, Okay, this is creative and this is not creative. But what

what it depends on is what the listeners thinks of this, right, um? So, which is then in a way makes it really difficult to do research because as there's no clear definition of what we're measuring. Um, it's it's all the subject driven. It's really hard to say, Okay, this is something where it's going in the right direction and this is not

so much. Yeah, but I mean it's so that the problem is kind of mentioned learning about official intelligence algorithms, they all try to they learned from data, and they essentially always try to reproduce something that they learned from the data. Right, while real creativity is always thinking all of the box. I wanted to be unexpected, like you know, uses these algorithms because he wants to surprise himself, but he likes to set certain conditions that are appealing to him.

It's sort of like being the prime mover in the situation and then just sort of letting the pieces fall where they may at the end of the day. But you are sort of still putting yourself into the equation. But then you are hoping for unexpected results to surprise yourself. And this is definitely one very good way of dealing with that. Right because you you have some kind of random components there, Um, you don't trust everything that is

being output. It right, So, but something might be good, So you generate a lot of variations of of what you might want to achieve, and then you just pick something that that really bokes and then you um use this as a starting point from where you want to

go to where you want to go. I mentioned elevator music, and I get that for sure, but aren't they already using generative music and video games where they have to have music constantly playing And obviously it would take ages for a single person to compose, you know, hundreds of hours of music. And I know there are cues in games that are composed, but then there are parts where you're maybe wandering around and like the you know RPG

type game and it's sort of ambient music. It just seems to morph and change, right, I mean, this is but this is rule based as far as I know.

I'm I'm far from being an expert in in what really happens in these game engines, but my understanding is that they define specific states, um and then they have certain rules for either looping something, looping specific loops or just generating some some more atmospheric background tones within a palet or within like a scale or something that's you know, but I'm I'm I'm pretty sure that this is not necessarily automatically generated. I mean, there might be randomness in there,

but I think it's basically rule based. So somebody during the development specified, okay in this state do something like this. How do you think that technology will shape music over the next ten twenty years. I mean, obviously, we're at a conference festival that is very much involved in the connection between technology and music. I love it. I think it's amazing. There are some people that are kind of freaks out, But I wonder what you think about, like,

where's it going? Oh, well, that's that's obviously very hard to answer. I mean, I mean, so, okay, let me start historically, right, so, so technology and music they have always interacted very closely. Right, So, there's actually genres who would not which would not be there without the technology technology exactly, so the electric guitar rock and roll wouldn't have happened with all the electric guitar, and the electric

guitar was in essence and engineering effort, right synthesizer. Obviously, we are here at mook Fest. Um so, so there was always close interaction between technology, um so. Um what the trends that I currently see, and they are not really surprising, I guess, but I think that, um, the interaction of the performer with any kind of sound generation of music generation will will um grow more cohesive. So any kind of controller will be easier to use and and uh, it will also be easier to use for

everybody to create music. And this is definitely a trend you already see with d J apps and so on, where they automatically create matchups for you and and all this stuff. Um, it's this is this is definitely going to happen that the user will be even if they have no music background, will be able to create music in a way that that makes sense. It might only be loop based for now, but there's a lot of possibilities here. Um, I see all the possibilities in more

crowd based approaches to this. Right, So, um, what happens if you put a hundred people into a room and give them, I don't know, an app or something that they can control and then they make music together? Neural network music exactly. And and there's also in this context

there's new forms how artists can communicate with their fans. Right, so you could release something that is actually interactive, So so fans could, in the easiest form could vote on something, but maybe some more complex input would shape the music and outcome there. So I think these are very very interesting forms where you already see the seats in what's currently happening, um, and I think this will definitely evolve.

Knowl and Lurch makes some great points about the subtleties of music and analysis as well as the potential for their future. And when we come back, we'll talk more about generating music from a computational standpoint. Generating music, like musical analysis, is a non trivial task. How do you program a computer so that it might dynamically create, esthetically leasing measures of music without becoming too repetitive or boring, or straying too far away from a melody line to

sound like anything other than just random series of notes. Now, some music, maybe even a lot of music, is written very deliberately. You know, your painstakingly sitting down and figuring out what chord comes next, when should you put in the key change, how many times should you repeat the chorus. It's not as if some mythical muse has reached down to touch the musician's brain and create the song fully formed.

But there have been attempts by humans to create music from an almost engineering perspective, so that it almost it almost feels like you're taking the artistry out. That's not entirely fair. I don't really believe that is so, but there do. There are some songs out there that were created by committee, and you could argue that some of them perhaps seem to have less merit to them than others. Now, there's some commit the design music that is amazing for

reasons that are difficult to put into words. For example, in n Dave Soldier, a composer, worked with two artists, Komar and Melamid, to create what they titled the Most Unwanted Song. They conducted a public survey to find out what people most liked and hated in music, and then they created two different songs that incorporated many of those elements. The ones that included the lowest scoring elements became part of the Most Unwanted Song. And it's a song that

lasts about twenty minutes. It's incredibly long. It's a song that includes accordion, bagpipes, children's voices, and opera singer rapping, and also incorporated advertising. It's gloriously awful and it sounds like this now. They also did the most Wanted Music, and they created a song that incorporated the elements that the survey takers identified as being the most pleasant components of music. The result is something that would likely put Kenny g into a coma. It's listening so easy you

don't even know you're listening. It's a shout out to Peter shik Ali right there. I actually think that this song is worse than the most Unwanted Song, but take a listen the world. Both examples illustrate the power of music analysis, as well as how it can easily be misinterpreted or misused, which can create I think we can all agree horrific results. But neither of those pieces were actually generated by computers. That was all the work of human beings. Human beings with a wonky sense of humor,

but still human. And you might think that the first computer generated music must have come a decade or so later. I mean, the Unwanted Song and Wanted Song both came out nine, but now was late for computer generated music. The first actual piece written by computer was the Iliac Sweet for String Quartet, created in nineteen fifty seven. This was the work of Learn Hiller, a composer, and Leonard Isaacson,

a mathematician, and their approach was fairly straightforward. They created a program that would generate pseudo random integers, which in turn would represent important information with regards to musical composition

such as pitch, rhythm, dynamics, and other factors. This processed information would then go through a pass on a filter, and that filter would force the data to follow rules of composition, so it sort out anything that went outside of the rules of composition and anything that was when then the rules would get a pass and the resulting piece of music for a string quartet sounds a bit experimental, but it doesn't exactly sound mechanical. It sounds kind of

like this. Other experiments and music generation followed, but they all depended pretty heavily on computers working within relatively strict sets of rules, with a good deal of human guidance along the way, and of course the computers had no actual understanding of music. You could program in rules for different musical genres and computers can do that. That's what

computers do. They're really good at following rules, but the machines have no way of knowing why those rules exist or what sort of effect those rules have on the music itself. Computer scientists have created some interesting experiments to build music generators. For example, Matt Vitelli and Erin Naiebe built software that analyzed a piece of music by Medean, a French DJ, from the day on I suppose I apologize my Francie is uh not very good. The software

analyzed Medeans work and then attempted to replicate it. It used recurrent neural networks an attempt to capture the essence of the music and make something similar. The neural network learned with every iteration of music uh, and learned how to more closely mimic the style, So when it first started it sounded like pure noise. It took two thousand iterations before it generated something that resembled a song more

than noise. But it shows that these learning algorithms are able to start focusing on what those elements are that represent meaningful information versus meaningless information. So would this eventually be able to create its own music if you were to say, said it to listening to a radio station for long enough. Who's to say? Over at Google, the Brain team is working on a ton of different projects related to machine learning and artificial intelligence, including exploring opportunities

for computer generated music. This falls under something called the Magenta Project, and the project has two purposes. The first is to experiment with machines creating different forms of art automatically, including music. The second purpose is to foster a community of artists and programmers to find new and interesting ways

to use this technology that Google has created. On the official page for Magenta, Douglas Eck points out that artists have always found innovative ways to put technology to use beyond what the creators had in mind, and that's where true innovation lies. So in other words, when you create an electric guitar for the first time, you're probably not anticipating the way Jimmy Hendrix is going to play that

decades later. So artists have been able to take things that people have created and move it beyond even the creator's expectations. That's kind of what they're hoping over at the Magenta Project. Ck goes on to point out that short form machine generated music can be quite effective, and it's been around for a while. There are generators out there that can make short songs essentially are short pieces

of music. But if you increase the duration requirement, if you require the music to last longer, you start running into the limitations of the technology. They start to become more apparent, and it becomes clear that machines aren't really good at sustaining a long term narrative in any format. The Magenta project isn't just a single approach. It's not like a group of folks who are just working on

one set of algorithms. Think of it more like a platform or a list of assets, a list of available uh bits and pieces other people can use, and programmers and musicians can build tools out of those pieces for generating music. Now, some of those tools may end up being way more effective than other tools. Just figuring out how to evaluate the abilities of the software could end up becoming a challenge. How can you tell if one autonomous music generator is quote unquote better than another one.

Music is pretty subjective and what I might like might not be what you like, And there are some qualitative elements that we can look at that are pretty difficult to to get a conversation going, because if you have a very different set of of pros and cons or or set of preferences I should say about music than I do. Then we might hit a wall. But there's some quantitative elements such as the amount of variation in a piece and whether the music generated fits whatever genre

you're aiming for, that you can use those. That's a little bit easier because it's a quantitative or more or less a quantitative element. But pretty soon you get into more subjective territory, and that's where it all breaks down. At the moment, machines are better at interpreting and combining

musical pieces than they aren't creating something entirely new. For example, David Cope, who is a professor emerit us at the University of California, Santa Cruz, is also a composer, launched a project called Experiments in Musical Intelligence many years ago and use the computer program to analyze various classical composers musical work. Then the program would construct new pieces using the elements it had analyzed as building blocks for that piece.

So the program wasn't really writing something entirely new, but rather combining found elements in new ways. Now, perhaps in the future machines will be able to make art on their own with minimal human input, and if that happens, we'll likely have to face some tough philosophical questions about the nature of art. If a machine doesn't possess self awareness or consciousness and really is just a complicated set of equations that generate data according to some general rules,

is its production actually art? Is intent required for it to be art? Does the artist have to intend something in order for it to be art? If people enjoy the work and find it intellectually or emotionally stimulating, does that make it real music? If if I like something and I find out later on that a computer generated it completely from start to finish, does that at all lesson the value of that music? Or does the fact that I like it mean that it's quote unquote real.

We're none at the stage right now where those questions need urgent answers, but I do think they're really interesting, and now it's time that we play our own music and get the heck out of here. So if you guys have any suggestions for future episodes of tech Stuff, right to me. Let me know what you think. Our email address is tech Stuff at how stuff works dot com, or you can drop me a line on Twitter or Facebook. The handle for the show at both of those is

tech Stuff hsw on Wednesdays and Friday's. I record in the studio and you can watch me live on twitch dot tv slash tech Stuff. Watch as I struggle for words and fail and then head desk and then tell Dylan to pause the recording so I can come up with something and then start the recording again. You get to see the whole thing, so all the stuff that gets cut out of the podcasts, you can watch it

happen live. Sometimes I dance. I hope to see you Wednesdays and Fridays at twitch dot tv slash text Stuff and I'll talk to you again really soon. For more on this and thousands of other topics, because it has stop works dot com