The Music Education Podcast
Listen for authentic and challenging conversation on all things music ed. Brought to you by Charanga. Hosted and produced by Chris Woods of The Chris Woods Groove Orchestra.
The Music Education Podcast
Episode 70 - 'A.I. In Music Education Pt1' - Gary Charles
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
In this episode Chris chats to Gary Charles about AI, music and music education. This episode is sponsored by Longy School of Music; find out more here... https://longy.edu/
Welcome to the Music Education Podcast. I'm Chris of the Chris Woods Groove Orchestra, and this podcast is brought to you by the Sandstorm Music Education Hub and our sponsor, Longe School of Music. In this episode, I chat with Gary Charles about AI, and this is part one of a two-part AI series. So, AI, where do we start really? I mean, it's a huge part of our lives without really even knowing it. And I think all of us are now conscious and maybe even just accepting of how it's going to probably become an increasingly large part of our lives. When it comes to music, I think things are a little bit more ambiguous. That's why I felt it was so important to make these episodes, really, to help explain some of the details around AI and to really look at its effect on music and crucially music education itself. Now, we've got two parts to this. The first part is very much an introduction as to what AI is. Now that might seem super basic to many people who think, oh yeah, I know what AI is, but I kind of thought that too, and actually Gary really breaks it down in this incredibly in-depth way, which really reveals some important pieces of information that I think can help shape how you think about it. So even if you're a bit of a techie, I would go with episode part one as well. Towards the end of part one, we start looking at some of the effects that AI has within music itself and start to look a little bit of how it might be playing a role in music education. And I posed Gary the very, very simplistic but I'd like to say powerful question of is AI actually relevant in music education? Is it important to know about? I quickly summarised the answer to that question as well, and his answer is yes. So continue to listen. When we get to part two, it's really a very much more in-depth and philosophical look at AI. And some of the things that Gary reveals are really incredibly fascinating and profound. We're not just talking about tech here, we're really talking about culture. Some of the things that AI reveal have an incredibly powerful knock-on effect on how we should view music education. Because what we're talking about is really the difference between human beings and machines. And when we think about it like that, those mechanistic musical ideas that we put into music education so often start to seem a little less relevant. I hope you enjoy listening. Gary, thank you so much for joining me this morning. Can we start off with uh who you are and what you do?
SPEAKER_02Uh excellent. My name's Gary Charles. I'm a uh artist, musician, um, researcher, and and educator. I've been doing a uh a PhD in in um music composition, dealing with artificial intelligence in in music comp composition for for a few years now, um, and getting towards the end, I hope. And um I I lecture mainly music production, but also things like uh some songwriting and uh cultural studies at BIM University in in Birmingham.
SPEAKER_01Wonderful, thank you so much. So the stimulus of the conversation is AI or artificial intelligence. So I think just to set up the conversation for all types of listener, could you give us uh a bit of a definition of what AI or artificial intelligence is in in principle or theory, even I suppose?
SPEAKER_02Okay. Um I mean that it it's sort of a bit difficult. Um so so for example, if you'd have asked me that um five years ago or six years ago when I when I sort of uh started research in the area, my answer would have been quite different because the the public consciousness is has has changed around it. Um so uh I think the the thing that I find quite useful is is making a few um making a few distinctions. So um weirdly today, I think potentially the best the best way to understand artificial intelligence as a term is almost like a marketing term. Uh so we've all um and I know people on the music side, if you're if you're a buyer of of plugins or any musical um material, you you're inundated with AI-based music tools.
SPEAKER_01And and I can without knowing it, right though, as well as it's not.
SPEAKER_02Without knowing it, sometimes sometimes with that being the the main headline of why you may want to purchase this tool, and uh without really knowing what that what that means. But the the the real kind of emergence of of of this term has come through um through things well based on generative AI. So things like uh chat GPT, which I'm sure sure people have have come across, that's that generates text, um, image generation tools like Mid Journey, uh DALI, and then there are also uh music generation tools, which we'll we'll we'll talk about. The key, the the key, I think, distinction there. So the way I like to think about it is it's a different type of computation. And the distinction between, because this is one of the things people often can't tell the difference between so what is AI and what is just really very powerful computers or very powerful piece of music or software. Um the real the distinction in its simplest form is that the the outputs of of the computation aren't based on on programmed information in in the the the sense that we understand normal computer programming. So a programmer um writes in symbolic code, if this happens, if XYZ, then the output is Y. In the case of of what has now become known as artificial intelligence, you'll see you'll see uh references to machine learning, um uh deep learning, all of these are based on um neural networks. So that this other type of computation that uh the way that it derives outputs is not on a simple programmed in basis. If this happens, then why? It takes a huge set of digital information, so it may be maybe text, uh digital images, digital um uh audio or but or digital sound, and all of that information is is categorized and labeled, and the neural networks are are designed to almost try and mimic the way that our brains work um in that neural neurons fire and connect with each other when they when they detect patterns or or uh yeah commonalities in in experience. I mean that the it's already getting maybe too complicated. But um so the the key difference is that then the outputs are determined by looking for patterns in a large body of information and deriving probabilities of what the output is based on on all of those those bits of information. So if we think about um chat GPT, which is the the the the kind of most known um of these these uh generative AI um type tools, it's it has a huge database of text mainly scraped from the internet, because that's where digital text is is based. Um and always for artificial intelligence uh models we need to digitize information. So that and that's something that um that gets us into the to some of the challenges with the with the technology. But um that that whole data set is scoured for for patterns. And so and that and that comes down to to quite simple things like the model looks at different parts of words, um strings of words, and it kind of creates a probability of what uh the next word will be in a in a string, what the next sentence will be in a string, and creates um new text based on patterns that that the model is detected in in a vast array of of previous text. Um and the same thing happens with with um images, so so um models like Mid Journey are um trained. So the training is is the part where uh this huge data set is analyzed for uh for common patterns and common labels, and um the outputs are derived from looking at the the patterns inherent in those in those huge data sets. So so that's from a technological point of view, that's generally what artificial intelligence is mainly looking at. But I I think on top of that, there's another really useful distinction, and that's this distinction between um narrow AI and and AGI. So um AGI is artificial general intelligence, and this is the the kind of um where the with a lot of the the the fears, the um kind of uh a lot of Hollywood movies kind of based around um a computer intelligence that has learned to generalize their um their knowledge derived from from historic data to be able to do things uh in in a kind of equivalent or superior way to humans. The bit that that um I think is more interesting is that the AGI and and um the company that that runs uh that created GPT uh OpenAI, they are expressly their their company motto is something along the lines of creating uh AGI for a better humanity or something like that. And I think it's worthwhile and unpicking what that means. Yeah. And so and and the interesting thing, I think, particularly for music folk, so when when um companies are talking about what AGI could be, certainly the the thing that that people go to first is is uh large language models like Chat GPT because language is such a uh an important part of our um of human culture and and stores of of human knowledge. One of the other um components that that people imagine as being what this really kind of advanced generalized intelligence would be is creativity. And the first thing that that um a lot of the general public but also technologists um consider as inherently creative is the creation of music. And so there the the large the the large tech companies who are developing um the software all have a music creation department. So essentially trying to create music and be creative using this this technology.
SPEAKER_00Is your music career ready for its next beat? Longy School of Music of Bard College offers a Master of Music in Music Education program online. Longi's top-ranked accredited program is rooted in culturally responsive teaching and designed with the working teacher in mind. Flexible options like asynchronous classes mean you don't have to take time off to pursue your degree. Longy School of Music, located in Boston, Massachusetts, is a leader in centering music as social change, expanding the world's understanding of what a life in music can mean. Our Master of Music and Music Education online alumni go on to do things like advance their teaching licenses, earn higher salaries, found community music organizations, and teach in higher ed. Financial aid is available, so apply today. Learn more at longi.edu slash study. That's l-o-n-g-y.edu slash study.
SPEAKER_01So why I love listening to podcasts and doing them is because uh yeah, so much of uh information that we consume is um in very short sentences, and you know, hence why things can be so divisive, and you've just proved the beauty of a podcast by um giving a a really in-depth sort of explanation of it, um which is interesting, right? Because I I think I've of myself as someone who maybe gets AI to some degree. I've listened, I've listened to a couple of podcasts about it, read a bit, sort of interact with it in some way, but everything that you were just saying was just making my head go, oh okay, do I actually get this? So can I can I try and explain it back as to how it sort of formed in my brain, and you can go Absolutely, yeah. You got it wrong, mate. Uh okay, so a c a calculator, for example, yeah would be the way that the computer or the calculator arrives at those answers, is it's been given a formula to work out those answers. So even though it's not been a calculator hasn't been given every possible answer, it's been given the specific formula. Yes. That isn't AI, that's just programming, correct? Yes. Yeah. So whereas AI would then be it's not been given a formula, it's been given pieces of information and instructions that then begin a process of piecing that information together. So the answers would be more abstract or answers to abstract questions.
SPEAKER_02Yes. Yes. Go on. Um so so the the the thing that's happening with with artificial intelligence, it's not calculating every single time you're asking it to do something. It it is a model based on training from a large so so um it's it's not always going back into the into the to the p massive pool of information that it's going into the model, which is essentially like a mathematical or statistical summary of all of that that um that massive pool of information. Because to train to train one of these models so to create that statistical uh summary takes a huge amount of time and a huge amount of uh computing power. Um so for example, um if if you if you go into to GPT and ask it about something that's happened very, very recently, it will have no concept of it. And in fact, they will they will say to you the let the last training update in the model is you know say the the 20th of January 2024. So even though this is a uh kind of sold to us as a very futuristic tool, it really can only look into the past. And so so the the thing that you're accessing is the model, which is the statistical summary, uh, or um uh somebody um Eric Silvaggio, who's a who who does a lot of work on um AI image generation calls calls it an infographic um of the database. So the the really important thing is is analysing what's in the data set. Uh and and so that's where the uniqueness of of the art artificial intelligence uh or what people kind of generally f refer to artificial intelligence. A really good ex a really good example from the music side is um so there are quite a few um there are quite a few tools that are that are advertised as AI um uh mixing tools, uh EQ tools. So so I use a um uh an AI assisted metering plugin um from a company called Sonable, where it will it will show you the average EQ curve of of sort of thousands of of um pop songs in a certain genre. Now that's something that we had before. So um I think um Wave's Q clone um was around for for ages, and you could kind of go and compare your song on the EQ curve to to a load of others. In in that case, in in with with Q clone, the difference is that the the coder has has coded in kind of the the bank of probabilities, or whereas with um the sonable plug-in, they would have trained an an algorithm on a whole range of of of songs that have been categorized, so they have different categories like electronic, um let's say drum and bass, and detected patterns in those in those EQ curves over time in in tracks, and then that is the the the band that you're comparing to. So it hasn't been essentially hard-coded in by by um a developer.
SPEAKER_01So is the machine then learnt in order to make it match this EQ curve. Um in my experience, if I turn the higher frequencies up by this amount, actually that sounds rubbish, so I need to do it thinks about all those different sort of outcomes, like an experienced engineer, I suppose, in that sense. So that's the sort of human element it's calling on its experiences of mastering a million tracks.
SPEAKER_02And with with with those type of very sort of very narrow uses or tools, um they definitely um are more there to augment or to to to make it easier for people to do do things.
SPEAKER_01So, Gary, can we have um a bit of a summary of some of those AI tools that are in the accessible musical landscape to uh people like you and I, and also to to young people often for free or for paying. We've talked a little bit about um mastering, I think. But yeah, what what what else is sort of out there being actively used?
SPEAKER_02So um in the sort of um uh tools to to assist music producers or musicians, music creators, um the technology's kind of been been used to to uh to create things like uh voice models. So voice uh voice models quite in the news at the moment because you can go and train a voice model on on um any voice you you really want, and and so we've had things like deep fake uh Drake songs, but some artists have been kind of at the forefront of that for a while. So there's an artist called Holly Herndon, who's who's quite um who's who's sort of the one of the first names that that will come up when you go and look at um the investigation of artificial intelligence in in um in music, and she uh trained her a voice model of her own voice, um, and it's called Holly Plus. And basically what what that is um is you can go and go and um sing something uh or or submit a WERE file. It's free to use, so she has a web uh a website that you can go and and submit a WERE file and then it sends it back to you, um sends it back to you in the voice of or in the voice of the the model of a Holly Herndon's uh voice. And it's still it's still quite rough depending on what kind of uh what kind of information you you sound it, you send it. But um to the the closest analogy really is um so they they they're thinking or they're they're putting that voice model in a guitar pedal, for example. So it acts like it can if you can get it to work in real time, then it acts like a a um uh an effect. So and already without AI we had convolet convolution technology, which is so if any if there are any um sort of uh logic users, there's a that you can either have an algorithmic reverb, which is once again programmed um programmed and and hard-coded, or a convolution reverb, which is tiny little samples taken of of of a space and and added to to the signal uh when it sends it through. So um and that and that and that's been around for you know for for quite a while. The difference once again then with with AI is is that the the model is generating those those pieces of audio based on samples that have been detected by the by uh by the training and that are in the training set.
SPEAKER_01And do you think to your if for a primary school music teacher, for example, teaching Yeah, like eight-year-old music, do you think there is any relevance at all? So they've just sat through the 30, 40 minutes of listening to this incredible knowledge from you. Do you think there's any point in them having any grasp of this stuff? Or do you think it, you know, for them and for their students it's novelty stuff? Or are we talking about tools that actually like you're talking about the voice generator there? You know, that surely can be uh uh as as a creative a tool as I don't know the cello is behind me, or do you or do you think it is just sort of not really for them?
SPEAKER_02Absolutely I mean absolutely. I mean I I would I would suggest that I would suggest that um in in I mean I'm not I'm not an expert on primary primary school or or or secondary school or GCSC music education. I've I've mainly um taught at higher education level. But from my own experience of of school music, um I would I would say that that very often these tools that that that students are being being exposed to may be much more relevant or much more exciting to them than than learning um you know notation or whatever the case may be. And and and I and I think even in higher education there's perhaps too much of a focus on sort of Western um Western musical traditions, and some of those are also baked into the technology that we use, but certainly school kids will be getting exposed to things like um you know deep fake uh MM rapping on vocals. So and and and it it it even applies to things like using DAW's or or um uh DJ techniques, so things that don't fall into the the the typical I guess piano centric or or or um typical uh music education at school. So so I th I think I think those and that's that's maybe not something that uh well I don't know what your typical primary education teacher wants to hear. I sort of feel that that the artificial intelligence thing, first thing in education that that everybody kind of really worried about was um was uh essay generation, um people submitting submitting um work that they've just asked GPT to generate. And in music education, I feel like that that's not so much of a concern. And in fact, that would only be a concern for the bits of music education that I think have been overcentered. So things like learning sort of uh Bach Beethoven, you know, is is the center of all music learning going going outwards, whereas um I think we need to take on you know kind of more diverse plural range of practices. Um part of that may be technology-based.
SPEAKER_01So are are we in a in a situation where specifically referencing music education, we're saying actually AI is an opportunity to bypass some of those kind of what people have traditionally in the Western music education world anyway, seen as the building blocks of music, and saying, I don't know, I don't know if you you need that because you could go, AI can offer this to you. That's always the thing, right? So I'd be I'd be so don't don't be careful, Gary.
SPEAKER_02Go on. No, no, I'm so AI the or as a concept, it's completely overhyped at the moment. So the things that the the the change that it brings, I would say, is far less than what um is perhaps being sold to us, particularly um in music. So most of the things that we've spoken about are really just sometimes improvements and and really impressive engineering, but they're kind of just um adding on to things that we were already doing and that were already there. If these tools actually encourage musicing and a more music as the social uh in the social practice, music is a verb as opposed to a noun, um, then I would see it as a as a really good thing. And there are there are some examples um of that. However, I would kind of go as far to say that that they solidify some of those those and and canonize things even more because the tool what what I would what I would encourage in education both at school and and in higher education is more of the composer or the artist mentality um trying to to generate culture or push things forward. Um and by definition these tools are based on um what's gone before. Uh and so so there's a there's a yeah, there's a tension there. So even even though these tools are are sold to us as futuristic and and really innovative, really they're just looking looking back and and reinforcing the way things have been done. So which which is fine, however, that they're sold like uh for example when Google ML, which is a uh Music ML, which is a uh music generation um AI, also free to use, so and it's web-based, so anybody can go in, type in play me a uh uh uh uh uh film soundtrack for uh a horror film um composed by by uh Beethoven, and it will generate something that you know is impressively not a million miles from from how you might have imagined imagined that. Um when it was released, a lot of the the sort of headlines were you know Google Google Music LMS has uh or ML uh has solved music. And it's kind of so these tools kind of validate music as only the output, only the the piece of digital audio that we have at the end. Whereas as educators, I think what we what we want to encourage is is and the things most of us love is that music as a social practice, as a as a uh an act of doing of of being, um and and so there's a there's a a contradiction between those two things. So so so while I I've spent a lot of time working with AI tools, and I'm always really keen and fascinated to to work with the latest uh technology that comes out, especially when it's hyped, like AI has. And it and it exists in in higher education as well, which is really still very based on on um sort of uh Western classical music, essentially essentially. And whereas most students' interaction with with music is very, very different. It's um you know uh sampling, you you may be uh DJing, you're hearing, you're hearing all kinds of of styles, genres. There's new music coming out of different different parts of the world that don't necessarily conform to to or are not easily read or understood through those through those conventions, but it's still it it's still music. And so I think there is a challenge to to making music education more uh relevant to to actually the reality of of of the world. There's a there's a brilliant um article by a uh composer and and academic George E. Lewis. I'm sure you you may have come across him before. He actually he is a jazz musician from the from the States and actually created a um an artificial intelligence agent to to improvise with him back in the I think the the the set the 80s. Um but he's he's uh published a a paper uh called Decolonizing New Music Education. And he talks about there's there's a line in there that that um I think is really useful, and I use it all the time with uh with our students, is that there's no such thing as a as a great composer. There's just a plurality of of practices, of interests, of and of ideas, and um that should be should be really encouraged. And he points out that art schools, for example, have gone further down down that route where whereas they would have been the equivalent of um you know uh life painting and oil oil painting as the the main practice of of fine art. Um now art schools are looking at multiples of different mediums, of different concepts, of ways of of teaching, of of music from of art from different eras, different parts of the world, different um concepts, whereas in music education, and weirdly, particularly in the the more elite uh universities, schools, we're we're still teaching the same music practices as as that's what music is. And I and I th and I think I'm not sure if I'm not sure if artificial intelligence challenges that, but certainly um music technology and and kind of the the ability to to um assimilate music practices from all over the world, so so communication and and and I guess the internet certainly pose a challenge to that, I I would say.
SPEAKER_01Brilliant. Thank you so much to our sponsor, Longy School of Music, and to Soundstorm for making this episode possible, and thank you to you for listening.