Is Artificial Intelligence Creative? Glenn Gould, pianist and technological visionary, and "Dear Glenn," the AI inspired by him - with Dr. Akira Maezawa of Yamaha

Transcript made by Otter.ai (limited editing)

 

Mira  0:00  
Welcome to Woman of Culture! I'm Mira T Sundara Rajan. Join me and my distinguished guests to discover untold stories from the world of culture.

Mira  0:59  
We met at the Grand Cafe of the Chinzan-so Hotel in Tokyo, which is set in an enormous historic garden visited by Basho, the acknowledged master of Haiku poetry, who lived in 17th century Edo as Tokyo was then known.

Mira  1:15  
I wanted to talk to Akira about Glenn Gould, the iconic Canadian pianist who made his last recording on a Yamaha piano in 1981. Revisiting the Goldberg Variations of Bach, which he had originally recorded for his debut album in 1955.

Mira  1:42  
When he re-recorded this work, Gould symbolically closed the circle of his life and work, a life spent in singular dedication to the music of Johann Sebastian Bach, with whom he will always be indelibly associated as his premier modern interpreter.

Mira  2:12  
Although his training, of course, was as a classical pianist, his artistic activities extended far beyond classical piano, and even classical music.

Mira  2:22  
He wrote extensively on the topics that interested him - music, technology, and history - and he made extraordinarily entertaining programs on music for the CBC, while also trying his hand at radio documentaries, creating the experimental "Idea of North" documentary and the more conventional, but no less interesting, biographical radio program on composer Arnold Schoenberg.

Mira  2:49  
Gould also played the works of Schoenberg, and of other serial composers from the early part of the 20th century, with extraordinary insight - although no music could, at least on the surface, be more different from Bach.

Mira  3:25  
This was an incredibly risky professional move, but the bet paid off, as he enjoyed a spectacular career as a recording artist without ever appearing in a concert hall after this time.

Mira  3:38  
Indeed, Gould was fascinated by technology, learning about the recording process and all of its technicalities, to a very high standard. And he went so far as to argue that the concert hall as a setting for music performance and listening would become obsolete by the year 2000. He argued that listeners could have a much better experience at home, listening to recordings and being able to mediate their musical experiences through the technologies that brought music into the home in the first place, increasingly allowing listeners to moderate the way the music sounded through acoustic settings and equipment.

Mira  4:18  
Gould was an incredibly deep thinker, and his fascination with technology took his reflections into surprising areas, not only including the question of whether concert performance would become obsolete, but also, yet more radically, whether the concert artist would also disappear.

Mira  4:38  
The technological innovation that triggered this reflection was the Moog synthesizer, developed by Robert Moog in 1964 and used by Wendy Carlos in 1967 to record the album "Switched-on Bach," which featured musical works by Bach played by the synthesizer.

Mira  4:58  
Gould proclaimed that this  was a remarkable musical achievement and a landmark in the interpretation of Bach

Mira  5:06  
Was Gould being a provocateur? Probably, but this assessment must be considered in context. He seems to have been, above all, an educator at heart, wanting to raise the level of listeners to meet his own, wishing for his audiences to be better informed, and to experience not only the joy of his artistry, but also the joy of learning. Provocation could serve this broader, ultimately ennobling goal. From a musical point of view, his statements about the Moog synthesizer are profoundly interesting.

Mira  5:42  
Was he really saying that a machine could best humans at that most human of all activities, music?

Mira  5:52  
September 25th, 2023, would have been Glenn Gould's 91st birthday.

Mira  5:59  
On this occasion, the Glenn Gould Foundation in Toronto held a special event on artificial intelligence, at which they released a newly designed artificial intelligence system called "Dear Glenn."

Mira  6:12  
Yamaha says that the project was "inspired by his unique creative style, and launched to explore the future of music through the use of artificial intelligence."

Mira  6:25  
Yamaha comments that "Glenn Gould was known for his devotion to recording with digital media, and an interest in rethinking the relationship between performer and audience.

Mira  6:36  
"The project to develop the system has been dubbed Dear Glenn as a tribute to the artists attitude, which was the inspiration for the idea behind the project. The work undertaken by the project is detailed by Yamaha as follows. Yamaha analyzed over 100 hours of Gould's performance recordings to develop an understanding of his playing style, and employ deep learning algorithms based on the data collected to create the AI system. In addition to ghouls audio recordings, AI learning data included human input in the form of performances by multiple pianists, who were admirers of Gould and intimately familiar with his performance style, raising the quality of reproduction to new levels."

Mira  7:21  
"Near instant performance analysis of fellow human players enables the AI to play predictively while interacting with human musicians, more than simply an automated performance. The AI reproduces the masterful touch of Glenn Gould to provide an inspiring and interactive experience of co creation between an AI pianist and human musicians."

Mira  7:53  
As a friend of mine commented, is this "dear" in the sense of expensive? It's the marketing of a human being as a commodity.

Mira  8:03  
When I had the chance to talk to Dr. Maezawa in Tokyo, I didn't hesitate to put my concerns forward. But I quickly learned something that made me feel differently. Akira himself is a musician. Our discussion was informed by his unmistakable love for music, and his profound respect for Glenn Gould's artistry.

Mira  8:27  
Here then, is Dr. Akira Maezawa. From Tokyo, with his take on "Dear Glenn," the AI based on Glenn Gould's inimitable piano playing, which Dr. Maezawa has helped Yamaha to memorialize, for better or for worse, in a truly unprecedented way.

Mira  9:00  
So yeah, so you said that you were keeping track of legal developments in Japan as compared to other places yet that already that's fascinating to me. So are you as a as a technologist interested in what's happening in the last GC that has been run? And so

Akira  9:16  
I am interested because, you know, our, my goal would be to deploy something for the, you know, our customers or people to use? Yeah, so, you know, it's very interesting for me, because it kind of, you know, one would have to think about some of these legal ramifications if there are legal ramifications, what can we do to make our to be able to deliver our, you know, my research access to the customer? So, it's up there practical problem for me

Mira  9:43  
Absolutely. Is your sense that the legal community understands well, the kinds of things that you're doing, especially let me preface that by saying like your work is particularly fascinating to me, because, as you said, I'm also a pianist, and I've done a lot of work on Glenn Gould. In fact, I'm not actually writing a book on clinical, and which is part of the reason I wanted to talk to you. And it's so interesting to me how what you're doing, it's kind of on the nexus between the technological development and also the creative side. So you have several communities that you're interacting with, you know, through your work. So yeah, so from that perspective, you know, what, what do you think about the legal?

Akira  10:24  
So I think that, you know, even though so as I said, No, I kind of go to conferences, you know, in music, tech, and even the researchers have very divided opinions, depending on their stance or what they think about, you know, for example, you know, how would I say, so, you know, some people might say, this is kind of, you know, sending a chap know what nobody, like the big corporations now have access to data, they have learned over the good models, now, they want to have restrictions so that other nodes whose job competitors from channels participating, or people might say, Oh, no, you're going to have new creativity and expand the creativity possibilities. So even in technological community, it's I think, you know, there isn't any consensus. And I think, you know, we are still kind of like, trying to figure out what is good and what is not good at knowing. And also in also in the artistic community, as well, I see lots of people who want to, you know, easily incorporate this as kind of like an enabler or to explore new possibilities of creation, whereas other people might be more reserved about, you know, the possibility of kind of having their creativity stolen. I think, you know, like, I think there's a leak in the recent judgment in the California regarding some civil suit against the jury. Oh,

Mira  11:37  
yes. Yes, actually, the attorney case was copyright registration. So they try to register this work that is essentially incorporating elements generated through Lea, you would know much better than me how that side of it works. But basically, she had images that were generated through AI and then text that they had written themselves. And so they tried to register the homework and the copyright office said, No, you could only register the parts that had human input. But that ended up being quite broad, because not only the text but also the overall compilation was deemed to be worthy of copyright protection.

Akira  12:28  
If there's a control inputs manipulated by humans, it's very difficult to identify like a boundary of what comprises this human inputs what compressed because No, this no a maybe until no prior systems, which know is rules that are clear cut is much easier. But with deep neural network, like these control inputs can propagate in very unexpected ways. Or since that can be foreseen, which is like a very great thing about, you know, I think the generation takes prompt, you know, propagates in unexpected ways to create the entire image, for example. So it's, I think it is, I can think, think from technological side, it will be very difficult as to find like a clear and non unambiguous way to define what constitutes creativity. Absolutely.

Mira  13:14  
Yeah, actually, I'm, I'm comforted in a way to hear you say that, because this is my sense as well, is that we're kind of barking up some of the wrong trees in terms of trying to regulate this. But I think the sense that I get is that in the artistic community, the stronger feeling right now is one fear, as you say, the creativity stolen in our CO opted, and also a lot of fear about being kind of replaced by machines. Which, again, I'd be curious to hear what you think. I mean, my own sense is that's more of a of a question of ethics, actually, in which is to approach this and there isn't anything inevitable about technology taking over? Yes.

Akira  13:54  
So it's like this whole LC issue, and that was like ethical, legal and societal issues on the east regarding technology, it's like, and I think it's half about creating, because, you know, whenever there's new technology to the vocal culture, there's always, you know, various forces involved. So I think, yes, that kind of I do not know how epic scope issues are resolved. But I think in general, from like, technical and technological perspective, we will have to know we really need to respect the artists and people who actually create these contents. No, it's an I think, you know, these no Glenwood projects or, you know, maybe, you know, we have no might have had very different a different ramped up again, for implications, you know, if we didn't know ask people for no, actually, no, as of no recording, listen at Glenwood, or maybe, you know, experts in Glenn Gould, know, to kind of get give us feedback on what you know, might have been useful. So this isn't anything ethical, but it's more like a personal personal to healing kind of thing to have a snuff defeat, have a feeling that this is something that if he were alive, might have been something interested in. And I think once you have that, no, then we can I think have that that's probably one way to deliver. And another very wrong way would be to leave scrape data from now in line and just know shuffling, the deep learning that model and we're much Oh, no. So I think that's No. So it's about know how to present and how to have mutual respect for different parties involved. At least I think that that's one way to start a mutual understanding and communication. Absolutely,

Mira  15:41  
I should, as I'm listening to you talk, I'm thinking that people like you, who are very creative on the technology side, actually have a lot more in common with the artists than perhaps some of the corporate entities. You know, not to speak about Yamaha. But I'm thinking about some of the technological companies and so on their outlook, Hollywood entertainment studios, their outlook tends to be a little bit different. But I like what you're saying in terms of connecting with the creativity on both sides. That mean, to me, that makes a lot of sense. And I want to ask you about the Glencoe project in particular, so I'm so curious, you said something very interesting, which is, you know, trying to try to understand what Glenn Gould himself might have thought about doing something like this, which of course, we can't do, because he's not here. But he does leave, you know, so much writing and, and so much discussion that he did about technology is enthusiasms one, so So can you tell me a little bit about how, how this project developed? How, how you think about Glenn Gould's place and all that because he is at the heart of it? Yes.  16:43  
So as the kind of uncomfortable talking as like a podcast about how he might have thought, because this is kind of like, you know, you can have like a strawman argument, you know, but I think this, you know, this, but the project develops as kind of like, so I have been working on the way to model human performance to better understand performance. And to understand performance is through training a generative model. So I was in the channel, feeding a music score, and tried to play the style of a human performance, and see what the model has learned. It's a very nice way to, you know, kind of, it has no practical applications. But also, it's kind of interesting, because the model, sometimes in the model creates very musically, you know, something that sounds very unnatural. But if you think about it musically, you can kind of find out why, you know, for example, maybe a system has overfit on like a phrase ending or trying to slow down the phrase ending, for example, then you would kind of hear something, no, I would say for example, if you train a model that's been entirely trained on Japan, and he moves out, then you would kind of sound as if Mozart has taken too much whiskey or something. But it kind of has a very interesting learning experience, because they say, so I kind of more want to, you know, create things that understands human performance, and does something useful for performance. So that was very interesting. But so then I was working on this, and another people and guys in marketing, want to know, they wanted to know have to kind of like deliver a message about how AI and music should coexist. And we have been doing kind of like a brainstorming. And we want to say, you know, this really applying this technology might be something useful. And we were saying, you know, which, you know, performer might be appropriate. And, you know, and I think you know, Glenn Gould came almost unanimously Firstly, because he has experimented a lot with recording technology, which I think even today, you know, in some conservative pianists might find uncomfortable salutely But I think no, he has been experimenting and exploring new creativity is almost like a DJ. And, you know, cannabis. No, he has been I know from what I think and I believe in that he has been quite influenced by Manola kind of like a McLuhan's media theories you know, and I think you know, it you know, it's kind of like a no making these no media communication one level more abstract through AI because it's still like a compressed model. It's just not communicating different via different channels. So I think there was this implication I thought was quite interesting. And, you know, also new Glenn Gould has a rather distinct musical scoring style so it's very you know, from technical development aspect is very good. It's very good to decide no unknown. It's easy to assess if the model is learning about his style or not. And fortino though he he played on our CFP and in his no like their career so that was I had no struck with the Yamaha sides. Yeah,

Mira  19:55  
the famous the last over variations, recording of course, wisdom. That's Very slowly. So it's so interesting, I'm just listening to you, I'm just finding it so fascinating

Mira  20:22  
there was one aspect of the decline that I was especially curious about what I was reading about it, it said that one of the things you were interested in was having it play music that Glenn Gould have not played during his lifetime. So that was fascinating. Why did you make that choice? Well, I

Akira  20:44  
I think it's more I think, you know, having these machine we learn human patterns would be would be interested, you know, to create something that for the next generation of music, musicians and music pieces that, you know, that came up after. Is that, so, you know, that's, that's really interesting and silly. So my initial idea, so no, I'm kind of work to kind of backtrack even more. So I've been working on this thing called Music Ensemble, automatic ensemble system, which No, I actually am, I played violin. So when I played the violin, I have A complement system that will play piano, for example, with me. So maybe, at home, I can play like a Brahms violin sonata or something, you know, have a piano accompany me. And I also always had this dissatisfaction about, you know, these, creating the music that when I play with a very skilled musical accompanies the accompaniment gives you lots of musical ideas and insights, it's not just having something that simply tracks you. And I wasn't trying to see what know, if I can create these musical insights, you know, that, you know, can provide this system, but also provide me with musical insights. Know, which getting got me interested into experts with modeling, and how you did not express and kind of how you can interact. So, so basically, I wanted to basically make a system so that I can I'll be able to play like, you know, maybe Beethoven by sweater was going to

Mira  22:21  
stick Yes, amazing. Actually, I

Mira  22:23  
can really relate to what you're saying as well, because I'm also a singer. So I sing in jazz, in addition to being a classical pianist, so I don't play jazz. And I would sing a lot more if I had a nice accompaniment that I could, you know, literally turn on and practice different improvisations and so on. So, you know, there are these, I don't know what they're called, but they're apps that allow you to play, you know, the structure of a jazz, it'll give you the harmonic changes. It's extremely unsophisticated. So imagine, you know, something like you're describing that would be absolutely amazing. So as a musician, I can completely relate to what you're saying. Or for pianist, even being able to play with an orchestral background and and practice your Concerto in these kinds of things would be amazing. So do you think that you achieve that with this technology? Or is

Akira  23:13  
that all it's long, there's a long way to go. So Medeco like, Well, for me, first of all, there are many aspects. One being that if you train, like you know, humans play differently in Seoul versus you know, in solo versus ensemble. So in our first premiere, and Ars Electronica, we also had, like a joint go to play notice this in the system play the pieces he has never played as a solo and also the chamber music. What turned out that no, during rehearsal is no play chamber music like a piano trio, and the piano duet with that budget from a Francesco tree sandal. And this is an AI, although, and what we found is that, you know, if you play piano duet or something, it's

Unknown Speaker  23:55  
it's it doesn't really stand that you can't really play with that kind of

Akira  24:04  
data, even if no, there's an adaptive mechanism to adapt to tempo or something. So it did that also kind of raises a whole new procyclical. You know, the strawman argument of Glenn Gould that have acceded to the other pianists always happy no could have happened would have gone this way. So that's no kind of like another no unanswered question.

Mira  24:23  
Yeah, I think that question is really difficult, because the whole point of chamber music is the dialogue amongst the different performers. And that does strike me as being once a uniquely human, but it strikes me as being essentially human even because the whole point of the chamber interaction is that the musicians talk to each other through music and respond to each other. So even that issue of like, you know, Glenn Gould would have done it this way, as opposed to, let's say, EVM and his way. I mean, they have to have common terms on which they're having the dialogue even if one performance so, but from your point Have you I think I can see also, if the technology were sophisticated enough that it would integrate a degree of this responsiveness to what's happening, which is just an amazing prospect. And I think very exciting when we consider the AI as a creative tool or creative partner. We don't talk enough about these things. Now what I hear about both, more more on the legal side, and anything else is the prospect of AI replacing musicians. Which is not what you're talking about.

Akira  25:31  
So I think, you know, usually, you know, I think in a when we know, researchers do research know, people know, I think, you know, music, music technology isn't, doesn't usually have bringing that much money. So people who do music technology research are people who genuinely love music. But I think, you know, there's always this note, corporate dynamics, and that kind of wants to know, and there's always this kind of thing. So I can definitely see why people might have these kind of reservations or concerns.

Akira  26:05  
So, I mean, I think, but one common thing might be, you know, it may be even though it's maybe like a regulation, but, you know, kind of try to find a way so that the musicians can actually benefit from these kind of mechanisms, so that it's not like one company and the provider of the AI that's kind of free riding on everything, or No, I think that might be sent. I think that's like one important aspect, though, that might know, like, can be controlled legally?

Akira  26:44  
No, I think, you know, even, but it goes, and then I was just want to make one note, I think, I think that might be interesting is that, I think, even with creativity as No, I might know, I learned I trained this model, I think, you know, with no, if I train or more, if I have no humongous data, I can maybe train more sophisticated models. So there's still some data limitations. But what I found is that, you know, it's very AI by definition learners, kind of like that little average. So everything was kind of reverts towards the mean. And I think, you know, I'm not sure how to, like know, one question, I'm having his his creativity, described by like, a mean, average, or is it like a deviation from there? Because I've noticed that, you know, for example, in like, a globalist record, you know, so the system learns kind of like a mannerism like, you know, oh, he might have kind of his to have short articulation or will have like a kind of like a slur at the first two patients or something, you know, or emphasizing their voices more in these kind of, but it's kind of like a more average, I guess, no description of Glenn Gould. But I think, you know, but I think what makes his real performance report is very special. I think it's some of the moments where it seemed like, though, he punches you in the face, or gray or something, you know, and there are these few, absolutely memorable moments, I think, in his recordings that signify, and it's not the touch, I think, but it's what these exceptions, and I'm wondering if that can be acquired through machine learning by learning some kind of average, you know, is it like something that's acquired? Or is there a need for some other model of creative process? So that's more kind of like, you know, machine learning, or at least, you know, I think there's kind of like a wait for the creativity, you know, for a no humans in no time. I think that that that that kind of no special moments are what I think humans should know, what's kind of focus more on creativity, and I think no AI should know. And it's maybe I can support some of the little liberal, I think, no, he mannerisms, I guess.

 

Mira  28:58  
Yeah, I so much appreciate what you're saying. Because one of the questions that I often hear asked is, what's special about human creativity? What can humans do that machines can do? And this question always makes me a bit uncomfortable, because I'm not sure that can be articulated in words. But I think you just did express it in words very well. You know, because when gold is not, for example, the some of his mannerisms or his style, there's, those are, those are more cosmetic, it's more how he chooses to deploy them. It's his will is his intention. It's his vision as an artist that brings those tools of clay and he deploys them in a certain way. As you as you explained it with your expertise. The machine is not set up to do that. It's set up to do something a little bit different. And the way you describe it again, it makes so much sense to me, I see a lot of complementarity between functions of the music and the functions of the musician. And I understand something I've been wondering about for a long time, which is that is AI a tool that can be used by Chris In a people incident good tool. And from your explanation, it

Transcribed by https://otter.ai