Artificial Intelligence (AI) in traditional Irish Music

This may be opening a can of worms, but, I will do it anyway. This evening while listening to The Rolling Wave program on RTE Radio 1, I learned of a project in which someone downloaded all of the tunes here on Session.org, and fed them to an AI (Artificial Intelligence) program and had it produce thousands of “traditional Irish tunes”. I have opinions (good and bad) which I will not go into just yet, but would be interested to hear others thoughts on the idea.

It’s an exercise in pointlessness.

Also, I listen to this program regularly and find it very much help ful in keeping in touch with things.

Have a look on this:
https://thesession.org/discussions/47876

This site was built using abc code notation, in other words taking notes from music and feeding it into a computer. The obvious following steps from there, unsurprisingly, could/would lead to composition using AI.

edit: see also ~ https://thesession.org/discussions/42712

AI is getting a lot of attention in the media and arts world right now, and peers of mine are always talking about the exploits.

The thing about Art is that when it is honest and sincere, it comes from human inspiration and emotion. Even when something is created without those foundations, there is a clear and apparent distinction from things that are created from those foundations, and that’s when coming from another human. There is no emotion, no feeling, no inspiration behind the actions of a computer. It’s called “Artificial” intelligence because no matter how deceptive that intelligence becomes, it will always be artificial.

A computer knowing thousands of tunes is like a person knowing thousands of words. Simply knowing words doesn’t give anyone the ability to write meaningful, eloquent, or even coherent poetry. Just the same, the common person knows thousands of songs, but that doesn’t give them the ability to write a lyric, melody line, or harmonic progression.

99% of the tunes will be crap even if they statistically fall into the pattern of ITM. 100% against the inclusion of AI material in the repertoire. These techniques have the capacity to flood sites such as this with huge volume of low quality.

The concept is academically (from a computer science perspective) interesting. Musically it is bereft. It is anathema to the concept of a traditional art.

If I was trying to write a tune and needed inspiration, a set of AI-generated tunes might be a useful reference set to look for a short phrase or a snippet of melody to work from.

Simply generating thousands of ITM-shaped note clusters might tickle the AI nerds as an exercise, but that sheer volume of material would then need human filtering to spot the one(s) that were actually worth playing.

And then that few would need to actually be played and adopted before it actually meant anything worthwhile, and/or qualified to be added to thesession.org.

I don’t know what makes a good tune, but I know it when I hear or play it. When AI can do *that* I’ll start to be remotely interested.

The only good thing about AI in Irish Trad is that even if AI writes a passable tune it can still be molded into something special by the people who play it and give it their interpretation. Many great tunes today are played differently from how they were originally envisaged. Some great tune settings today are the work of many musicians over many years.
I would think that AI would work very well for Movie background music and elevator music 😉

Actually, the majority of human compositions are inferior too but are either ignored or even don’t see the light of day. The composers would likely bin them themselves before that.

The problem is that AI tunes are more likely to be inflicted upon us with less “natural” selection or censorship.

The “natural” selection or censorship can be a tricky one though, as a tune needs to go through the test of time before it can be deemed good or bad, I think. Sometimes, upon listening to a tune for the first time, I’ll kinda dismiss it as “nothing special”. But then there ends up being some little phrase or haunting element that later gets stuck in my head and brings me back to the tune. “Where was that from?”. Similarly, I guess a tune that seems good upon the first listen could end up losing its magic after a few times.

Well, then let’s hear or see what the AI has created from all of the session.org tunes. I’m curious!

I heard about the project you were talking about and wasn’t very impressed with the tunes it came out with, as they lacked consistency.

However I recently experimented with asking ChatGPT to write me a jig in ABC notation and was impressed by the output, being of similar quality of something I could produce myself.

I was even able to ask it to refine certain points, giving it a more melancholic feel (it was able to understand what I meant by this and the music reflected it). At the end it was even able to suggest some titles that were appropriate for the feel of the jig.

Of course on one hand it’s lacking any history, but I feel that if I started playing it regularly then it would take on something special simply through that process, that I’d be able to put the human aspect into it.

It’s an interesting can of worms that has been opened, though the process already began long ago when the tradition became less oral and opened up to written notation. I think it’s interesting for example that some typos in collections from the like of O’Niel became standard, that the human aspect of the tunes is already open to the caprices of technology.

Of course many (most?) will always dismiss these tunes, but I’m curious to see what future generations will feel about it all. In a way using a tool like this isn’t all that different to learning the norms of music theory and then applying them to your own compositions.

Asked for another one to share as an example, this is with no feedback from myself:

X:1
T:Days Past
M:6/8
L:1/8
K:G minor
G|:d2d BAG|A2A ABc|d2d BAG|F2F FGA|
d2d BAG|A2A ABc|d2d BAG|1 F2F G2A:|2 F2F G2f||
|:g2g gfg|a2a agf|g2g gfg|e2e edc|
B2B BAG|A2A ABc|d2d BAG|1 F2F G2f:|2 F2F G2A||

Finally - a solution to the problem of the dearth of traditional Irish tunes! I mean, who among us has never said to himself, “So much time, and so few tunes”?

1) this is not new/news, people have been writing tune generators that use the thesession.org database for years now.
2) who is going to play AI generated tunes? (a) you can pay hired-gun musicians, (b) you can play them yourself, (c) have your midi synch play them. the ultimate muzak, coming to an elevator near you.
3) database-based generator will not write “new tunes”. by definition, it will generate rehashes of tunes in the database. 2 bars of star of munster, 3 bars of the kesh, 1 bar of catharsis, great new tune! if you want truely new tunes, “monkey with typewriter” is a better way to go. a “good tune/bad tune” filter is needed in any case.
4) lawyers heaven. if you want to commercially record an AI generated tune, you need copyright permissions. You do not own the copyright (you did not write the tune), the AI company does not own the copyright (they “stole” the thesession.org database), thesession.org does not own the copyright (individual contributors/composers do).
5) as our morris squire said. why do we need new tunes/dances, there is already so many excellent old tunes/dances!

I have to say I find the AI tunes a lot less interesting than the reaction they generate.

Tune creation by human: idea (heavily influenced by knowing and playing loads of tunes), adding, fiddling, removing, adding something new, more fiddling, repeat 200x, play it for ppl, go back to add and fiddle, repeat, finally maybe stop fiddling and name, play over and over, hope ppl like/learn/share, maybe record, tell stories about.

Tune creation by AI: same as for human, but the initial idea comes from the AI (which like a human has “looked at” loads of tunes. 90% of turning idea into tune is the same process.

Chris

Chris, this is not how generative and database driven “AI” works. It does not start with generating an idea then elaborating on it. That’s the main reason people say there is no “intelligence” in “AI”. The word “intelligence” in “AI” is like the word “silver” in “German silver”, to establish which real thing is being replaced by a cheaper substitute. (hint: there is no silver in german silver).

Mando cello--

AI looks at a million tunes and generates an idea from those. You or I listen to a million tunes, and generate an idea from those. AI has to infer the properties of Irish music, eg scales, rhythms and other patterns and then away it goes.

Chris, your sense of how trad tunes are created strikes me as missing some important dimensions.

I’ve written some trad-style tunes, but they weren’t generated from the other trad tunes I know and play. My own tunes were inspired, usually by a person, place, or event, and by the emotion(s) I felt.

AI can’t do that.

At this point, AI also cannot physically play fiddle and invent new music from experiencing the joy of a bow hand and lefthand fingers dancing on the strings.

Gimpy--

When I listen to the tune you created, I’m not hearing your emotions. I’m hearing a tune. I mean, obviously it might connote certain emotions in me (and these might be the same as yours). I also have no idea (unless you tell me) what inspired your tune. Again, I’m hearing a tune.

Your tunes weren’t generated by the other tunes that you play…but their parameters sure were. You wrote in6/6 or 4/4, you didn’t put D# in a dmix tune, etc, and phrasing etc all influenced you.

Chris, I was responding to your comment about how tunes are *created,* not whether tunes accurate express and convey emotions between people.

Certainly, when a new tune is created within a genre or style, it’s informed by what the composer knows about that genre or style. But that doesn’t mean such information is the germ, the inspiration, for the tune. Your earlier post made it seem like that’s all that goes into a tune—parameters of the style. No, sorry, it goes deeper than that.

In fact, a person might very well put unusual accidentals into a new melody, see Paddy Fahy. One of my own tunes does precisely that, in several places, and with not just one odd pitch but several. Or a measure of 6/8 feel in a reel (Frank’s or Trip to Miriam’s). Such creativity today’s AI likely would reject.

To be fair, a lot of modern-day human-composed tunes DO sound like nothing more than intellectual exercises, combinations of notes rather than musical or emotional ideas. Such tunes leave me cold.

Regardless, I would argue that music does in fact convey emotions, often better than other modes of communication. The tune *is* the emotion. Of course, that’s as much about how the tune is played, but that’s the ultimate expression of music—not dots, not an idea in the composer’s head, but the physical act of moving sound waves that find an eardrum. Abstract? Certainly. Many emotions are ineffable. And that’s why a concert hall full of people can experience very similar emotions when, say, The Gloaming plays the Sailor’s Bonnet set, or people can’t help but dance to Dervish sailing through Mick O’Connor’s/Jack Roe/Palmer’s Gate.

some composers are very successful in writing tunes that evoke emotions and images. tunes like Catharsis, Ashokan Farewell, Far Away, the Butterfly.

I guess AI could do the same. “HAL, please write me a happy irish air. Dave, I’m sorry, I’m afraid I can’t do that, irish airs can only be sad. how about about a happy polka instead?”

I asked ChatGPT “What are the possibilities for Artificial Intelligence (AI) in traditional Irish Music?” Here is the reply I got:

------------------------------------------------------------------------------------------------------------------------

Artificial intelligence (AI) can be used in many ways to enhance and support traditional Irish music. Here are some possibilities:

Composition: AI can be used to generate new pieces of traditional Irish music by analyzing existing tunes and creating new variations based on patterns and structures that it learns. This can help to create new music that is faithful to the traditional style while also exploring new avenues of creativity.

Arrangement: AI can also be used to arrange traditional Irish tunes for different instruments or ensembles. This can help to create new versions of traditional tunes that are tailored to specific instruments or performance contexts.

Performance: AI can be used to support live performances of traditional Irish music. For example, AI can be used to analyze the sound of an instrument in real-time and provide feedback to the performer, helping them to improve their technique and performance.

Education: AI can be used to support traditional Irish music education. For example, AI can be used to provide personalized feedback to students, helping them to improve their playing and understanding of the music.

Preservation: AI can also be used to preserve traditional Irish music by analyzing existing recordings and scores and creating digital archives that can be accessed by future generations.

Overall, AI can provide a valuable tool for exploring, preserving, and enhancing traditional Irish music, while also opening up new avenues for creativity and innovation.

------------------------------------------------------------------------------------------------------------------------

An answer, I think, which is like gnawing on rubber. I wonder if the music AI produces is as tedious and shallow.

Some interesting and thought provoking discussion here, which is what I hoped for when I started this discussion, though on hindsight I think the “can of worms” was opened by the developers of ChatGPT before I made my post.

My own thoughts are these. I have written a number of tunes myself, some good (my biased opinion but also the opinion of others that have heard them), some not so good. What I think is important is that the good ones always came from some emotion/feeling I had at the time. The not so good, some downright bad (which no one else will hear) come from the idea of simply wanting to write a tune and just trying to put one together using the traditional forms, which is essentailly what AI would be doing. I feel that the emotion, the feeling must be there and this is something that AI can not do.

As for reviews, opinions, that some of these AI produced tunes are “good” and “worthwhile”, one must keep in mind that they are just opinions, no matter who the “expert” may be issuing them, and everyone (me included) has their natural cognitive biases, one of which is the tendency to only see patterns that confirm what they already believe. Thus if a reviewer is prone to wanting to see AI produced music as worthwhile, they are going to “find” justification for that (and the reverse). All of which means just because you get some reviewers to say it is worthwhile, does not make it so.

Some also will say “What’s the harm? It it’s simply another way of generating ideas and should not be looked down on.” Well there is a very real danger here, that of AI technology being used to quickly generate massive amounts of material that overwhelms the material that is produced by human’s based on feelings and emotion. If you think this is not a real threat and that the good stuff will naturally rise above everything else by virtue of getting repeated, consider this. Look at the state of pop/rock music today. Pop/Rock music a virtual wasteland of creativity because it is all generated by individuals seeking to reproduce the latest “hits” with some minor twist. This is not just my opinion, a few years ago I read a study done with a computer analysis of all a massive amount of the pop music and found that most of it (I don’t recall the details now) was virtually the same intervals, speeded up slowed down, accented differently etc, and writeen by the same relatively small number of of individuals. The point being the so called creators are just cranking stuff out based on existing formula of what sells, much the same way AI creates traditional music. Furthermore, effects of AI generated artistic material is already being felt. Clarke’s World, a SCI FI/Fantasy magazine has had to close submissions due to being overwhelmed with AI generated material since the release of ChatGPT. At this time they have no real answers as to how to deal with the issue and jeapordizing their policy of encouraging new writers and writers from diverse parts of the world. I find this disturbing and I see a real danger of this happening here and elsewhere.

In another discussion on a tech forum on the subject, one comment offered the opinion that “2030 will be the year cyberspace becomes uninhabitable”. I think he or she was being optimistic, things like this could make it uninhabitable much sooner.

A computer program can’t create meaningful art because a computer program doesn’t have an unconscious.

Jon Corelis, what do you mean by “an unconscious”?

‘AI’ is just a statistical model, it ‘learns’ the general patterns in the data (in this case, which note is probable given ‘set x’ of previous notes - then repeat. It isn’t really intelligent.

ITM and other art forms do evolve over time, and I don’t think AI is capable of doing this as its an inherently social thing dependent on the willingness of the players in the tradition to adapt new ideas into the music.

To get an idea how this stuff works, Computerphile made a video discussing stable diffusion, and ‘Let’s build GPT: from scratch, in code, spelled out’, explains how ChatGPT works.

Robert, I think your comment “It isn’t really intelligent” says it all. Calling this stuff “intelligence” is a misnomer to start with.

Pastiche?

I also think that generating tunes, outside educational drilling (sight reading or ear training), is rather pointless.

For me, ITM has always been about the social aspect of playing in sessions with others. I prettymuch only learn tunes I’ve herd often in sessions.

Tunes don’t have much value unless a considerable number of people are playing them.

Since no one has answered Jerone’s question, I’ll give it a shot. I believe what was meant by “an unconscious” was “an unconscious mind”; i.e., a “subconscious” - the idea being that no “meaningful art” can be created without some input from the “subconscious”.

Robert, I agree with your point that human social interaction is fundamental to this (and all) music, and since AI isn’t one of us, it’s tunes are pointless, nothing more than empty exercises.

And I get the drift of what you mean when you say, “Tunes don’t have much value unless a considerable number of people are playing them.” But that can’t be true at the outset.

Surely all the tunes we play started off being played by just one person, or maybe two in the case of co-written tunes. Paddy Fahy’s tunes had an audience of one at first, and I’d suppose that they had significant value to Mr. Fahy himself or he wouldn’t have bothered creating and playing them. Same would hold true for the compositions of Paddy O’Brien, Sean Ryan, Vincent Broderick, and so on. A tune like Splendid Isolation no doubt was first played in isolation, until it caught on.

I also know and play some tunes written by a local tunesmith. No more than a handful of people play these tunes, but they are powerful, meaningful tunes to us. I’d wager that’s genuine “value,” and who’s to say those tunes might spread until 10 or 100 years from now they’re played in sessions all round the world.

For me, a big part of why a tune has value is that it comes from someone who’s lived in the tradition, or in some neighboring tradition, and so is informed by that life. And that value holds whether a thousand people or just one or two play the tune. That’s what AI can’t do, and so far at least, the AI tunes I’ve seen or heard are just sequences of notes, not living music.

Never mind that across generations tunes wax and wane and wax again in popularity.

Actually I think there are AI tunes which have been played by people in sessions. Or am I mistaken?

I’m sure you’re correct AB. Humans are very prone to serious errors of judgement.

> ‘AI’ is just a statistical model, it ‘learns’ the general patterns in the data (in this case, which note is probable given ‘set x’ of previous notes - then repeat. It isn’t really intelligent.

This is commonly said, but’s a bit analagous to saying a cup of tea isn’t ‘hot’, it’s just vibrating atoms. Yes it’s a statistical model designed to predict characters, but also yes it has genuine intelligence… by any sensible definition of intelligence at least.

It turns out to generate really good statistical predictions of characters, these AIs, generate further and further abstractions which leads to them basically developing genuine intelligence and reasoning abilities - and whats amazing is that this intelligence emerges from a model designed just to predict characters. It’s not a human intelligence, it lacks the direct experience of being a human, but it can argue and reason about things in an intelligent way, (if they are written about extensively, or explained in enough detail).

But I predict whilst it might be able to create ideas, this iteration of AI language models will not be able to create truly decent Irish tunes, except by random accident - I would reason as follows:

- The input for the AI model is limited to the abc notation of tunes
- The abc notation does not capture the way a tune should be performed and whence how it should sound, of which there is some rough unwriten consensus between listeners and players of Irish music.
- There is no dataset capturing this notation->performance->sound correspondance for an AI to learn from.
- Because, the AI doesn’t know how the tunes sound, it cannot learn about the internal logic in the tunes, how they correspond to feelings, how they break expectations, how they resolve themselves etc
- Furthermore, humans compose tunes in this realm of abstract sound and (unconscious / spontaneous) feeling rather than notation.
- Because the AI has no access to to the sound and feeling content of the music it can’t ever ‘reason’ about how a new tune should go.
- Because it cannot ‘reason’ about musical ideas, it cannot create new ones (e.g. by inspiration) except by accident.
- Similarly because it cannot ‘reason’ about music ideas, it cannot tell which accidents are good.

So it may be able to produce tunes which sound kind of like Irish, and everynow and then one might be ‘goodish’, but there will be an element of chance and then of curation from a human, so it’s -cheating-.

If you sat a prisoner down with abcs of all irish tunes, and said ‘write 100 more of these’, there’s bound to be one or two that sound passable by accident, same with haikus etc.

The intelligence of all these AI models is limited to the datasets they are trained on, and for an AI to stand a chance at composing decent irish tunes, it needs to be trained on audio. The question would then be, whether an AI, seeking to predict the next ms of audio after being trained on all recordings, could plausibly generate a good performance of an Irish tune - which I am less sure would not be possible…

Cheers, I’ll take that as an assumption, bogman. Witty, of course.

AB, I was referring to those who might choose to play an AI tune, in case of any doubt!

I know you are, bogman. Cheers!

A friend shot me a link to a recent op ed about AI by Ezra Klein in the NY Times: https://www.nytimes.com/by/ezra-klein

(The link may not work for everyone, or the NY Times’ paywall may block many.)

I’m posting it here not to be alarmist, or to endorse Klein’s thoughts about AI. More as food for thought, as AI pertains to traditional music, sessions, and culture.

I don’t know Klein’s credentials to talk about AI, but he claims to be spending lots of time talking with people who are in the thick of the race.

One of his central points is that AI systems are crunching such vast amounts of inputs and running so many layered algorithms that no one can possibly understand in any detail what AI is doing. Even the developers feeding and running these systems say that—already at this point—an useful explanation would be too information-dense and complex to grasp in any one person’s lifetime.

And then Klein writes this, which I suspect is the piece most relevant to trad music:

“That is perhaps the weirdest thing about what we are building: The [AI’s] ‘thinking,’ for lack of a better word, is utterly inhuman, but we have trained it to present as deeply human. And the more inhuman the systems get — the more billions of connections they draw and layers and parameters and nodes and computing power they acquire — the more human they seem to us.”

It’s highly unlikely that we’ll put the brakes on this process. It’s careening along at a frenetic pace, fueled by
corporate and national desire to “win” the race and by oceans of money, with no regulation and no one coachman holding the reins.

Klein’s take-home point is that we humans have at best years (maybe months) to adapt to this, to sort out how to live with the blurring of what’s human and what’s machine code, and the many ripple effects that are already changing all aspects of society, of life.

He compares it to the early days of the pandemic, when most people thought it would quickly blow over, that tomorrow would be pretty much the same as today, and the few folks who dared think aloud about wearing masks or working from home or prolonged disruptions to our way of life were labeled as loony or worse.

I don’t know where we’re headed, but I’m glad I know a large store of tunes that assuredly were created by human beings, and that I have a few friends eager to play those tunes over a glass of fermented plant-derived molecules in liquid suspension. That will do ’til the end of my days, however late or soon that comes.

Sorry, the link to the op ed should be:
https://www.nytimes.com/2023/03/12/opinion/chatbots-artificial-intelligence-future-weirdness.html