London workshop on computer-generated session music

Re: London workshop on computer-generated session music

Sounds very interesting, especially in the light of recent discussions here.
Shouldn’t this (also) be posted to the Events page - https://thesession.org/events ?

Re: London workshop on computer-generated session music

Thanks Rick. I have added it to the Events.

Re: London workshop on computer-generated session music

Entering thousands of ITM tunes into a computer and having the computer generate music that follows the same patterns and tropes isn’t really a big deal. You could do the same by taking any twenty reels (in the same key) and randomly combining two measure phrases. Using reels in different keys would generate "more interesting" tunes.

Re: London workshop on computer-generated session music

Computers are good at taking the humanity out of basic endeavors and leaving us to simply choose the right button. Computer-generated "music" may be entertaining to some, but the result will be predictable.

Posted by .

Re: London workshop on computer-generated session music

Thanks for your comments.

Ergo: I am happy to see you are still around and maintain consistency in your skepticism of computers. Here is a video of me playing a tune generated by our system: https://www.youtube.com/watch?v=omHhyVD3PD8 I find it entertaining and unpredictable, just like the title! 🙂 But I am interested to hear you play the tune. Is it so far from humanity that you find the tune just cannot be saved? What changes would you make to the tune to help it transcend its mechanistic origins? Anyhow, if you are in London and willing, it would be fantastic to have you among the workshop attendees!


David: In a previous thread (https://thesession.org/discussions/39604) I try to make clear where our interests lie. In particular: "What we are trying to do is to see how the conventions in session music might be modeled statistically through the ABC transcriptions provided by practitioners on thesession.org. How do we know the ways a given model succeeds and fails? One way is to have the model generate new transcriptions and then compare their statistics to those of the real transcriptions. This offers an elementary sanity check, but has limited usefulness because it is disconnected from real world practice. Another way is to have the model generate new transcriptions and ask real practitioners in what ways those transcriptions succeed and fail. This is much more useful for shedding light on what a model has actually learned, and how we can "improve" it. Why even do this in the first place? … One reason is that the cultural sphere remains a exciting frontier for integrating computer technology. Most people probably don’t know the true extent to which computers and software are already integrated with the production of culture — the entire pipeline is digital. But we have so far to go, beyond the Spotifys and Netflixs. What’s the use? One possible use of this can be in a music transcription scenario: you are at a session, you want to see the dots of the group playing, so the music transcription system listens and works with the knowledge we have given it of the tradition to produce the dots, the name of the tune, its history and variations, etc. It needn’t be session music, but that is where we are starting for many reasons."

The approach you mention, recombining existing material into new material, requires only a shallow understanding and is not very interesting or novel. Can we make a machine do better then? How deep can its understanding go? Can it learn "for itself" a set of relevant rules/conventions that it can apply in novel situations? Can we have it springboard from an understanding of these rules/conventions to a new domain, such as atonal music? (Yes we can! https://highnoongmt.wordpress.com/2016/12/24/taking-a-christmas-carol-toward-the-dodecaphonic-by-derp-learning/)

Re: London workshop on computer-generated session music

I’m hoping to come along to the event.
At the risk of sounding like I’m doing our music down, I do suspect that folk tunes shouldn’t really be much of a challenge for artificial intelligence: for the simple reason that there are a thousands of them to work with (so a vast amount of data to start with); and also a lot of formulaic ones among them. Frankly, there are quite a lot of tunes that I encounter that were of course written by a human being, but which may as well have been written by a computer! I can’t be the only player that flicks through a book such as O’Neill thinking ‘next… next… next… next…. next’ before they hit a good one.

And then again, at the same time, there’s plenty of tunes that I’ve dismissed like those ‘nexts’ which I will then completely re-evaluate when I suddenly hear a good player doing something interesting with it. So there’s a sense in which it actually ought to be a win-win situation for a computer generating Irish tunes: even it can only come up with something formulaic, so long as the notes make some kind of melodic sense a good player will still be able to make something of it.

What I would be more interested in seeing is what a computer would do with all the tunes that I have singled over the years as being sui generis: the ‘odd’ tunes. If you fed nothing but 500 of the most idiosyncratic, eccentric Irish tunes into an algorithm: now that I’d be interested in hearing.

Re: London workshop on computer-generated session music

Thanks for your comments Matt! That is a risky position you take because it seems so automatic in our Western culture to equate value with complexity. And, "If a computer can do it, then what’s the use for a human to do it?"

One musician we are working with mentioned that some tunes in our generated collection pop out at him because of their familiar oddity… like some are written by someone who didn’t understand music theory, but liked the sound anyway. Other tunes have completely novel aspects but still work in the genre of ITM: it’s like the computer is helping us explore places in that space that no one had thought of yet.

Looking forward to meeting you!

Re: London workshop on computer-generated session music

Evolution is a march towards increased complexity, life is signified different from the inert by similar measure. Computers may model and imitate but they will not be the tradition in traditional music for many years to come. Sticking to cash generating algorithms that recognise and generate scores from monitored performance sounds much more appealing and useful to this sceptic.

Re: London workshop on computer-generated session music

"Is it so far from humanity that you find the tune just cannot be saved?"

Regarding that YouTube clip, I think the main problem is that it doesn’t sound like a tune anyone would want to to dance to, with that particular broken rhythm. Dance tunes are what we (mostly) play here on thesession.org, even if it’s no longer for actual dancers in most sessions. And some of us do actually play for dancers on occasion.

Maybe the AI should make more progress in understanding the dance rhythms, before it gets around to writing melodic lines?

Re: London workshop on computer-generated session music

Steve: Thank you for your comments. Leaving aside interesting discussions about teleology, complexity, and agency, I am wonding what you mean by "sticking to cash generating algorithms". Could you elaborate?

Conical bore: Thanks for your observations. "The Drunken Pint" is an odd tune, which we hand picked from 70,000 tunes generated by the first iteration of our system — the title is what caught my eye at first. It is not a good reflection of the kinds of tunes our system typically generates. To get a better idea, you could page through Vol. 1 (of 10) of generated transcriptions: https://highnoongmt.wordpress.com/2016/09/12/folk-rnn-session-tunes-volume-1-of-10/

Anyhow, I agree that our system operates in an "artificial realm" divorced of the _purpose_ of this kind of music. The only thing it knows is the ABC alphabet, and how to arrange it into sequences that are highly probable given the 23,000+ transcriptions of thesession.org. It knows nothing about the use of the music.

Re: London workshop on computer-generated session music

Hi all, what about the dancability of this computer-generated tune?

X:0
T: The Glas Herry Comment
M: 6/8
L: 1/8
K: Gmaj
|: G2 G BGB | dgd B2 G | FGA dGF | FAF DGA |
G2 G BGB | dgd B2 G | ABc gec | dBA G3 :|
|: f2 d A2 f | f2 d ABc | d2 d Bcd | ecA A2 g |
f2 A A2 d | eed B2 c | d2 B ABA | G3 G3 :|

https://youtu.be/QZh0WSjFFDs

Re: London workshop on computer-generated session music

Here’s another, but I play it more like a march than a jig. (Can it be played like a jig?)

M:6/8
K:Dmix
|A2D D2B|AGE c2A|GEE c2E|E2D DEG|
A2D D2B|AGE c2d|ecA GEA|D3 D3:|
Add Add|ede dcA|GEE cEE|GAB cBc|
Add Add|edc AGA|BcB AGE|D3 D3:|

https://youtu.be/eoySGZ5iJ00

Re: London workshop on computer-generated session music

"Ergo: I am happy to see you are still around and maintain consistency in your skepticism of computers. Here is a video of me playing a tune generated by our system: https://www.youtube.com/watch?v=omHhyVD3PD8 I find it entertaining and unpredictable, just like the title! 🙂 But I am interested to hear you play the tune. Is it so far from humanity that you find the tune just cannot be saved?"


Bob: I see that you’re still around and as wonderfully naive as ever about computers.

I find nothing likable about this thing you played. On paper it looks like a tune, and that’s part of my argument against polluting the air with a bunch of new stuff. Why does it exist? Because you thought up a cool new use for AI? I’d like this thing better if you included the words "computer-generated" in the title. Or better yet, your own name. I don’t see that anywhere on the page.

Hearing this, I would have to ask, Whoever wrote this tune — why? There’s no musical point to it.

Thanks much for putting these things out on Youtube, unidentified and unaccountable.

Posted by .

Re: London workshop on computer-generated session music

Thanks for your perspective Ergo. You are a very strong defender of the tradition. I am sure not everyone here at the session.org maintains such an absolute position. The professional musicians with whom we are working are quite open to these ideas.

Every one of my YouTube posts of me or someone else playing these tunes says something to the effect of, "I perform a piece compsed by folk-rnn — a recurrent neural network trained on transcriptions of session music." Also, the title of the workshop next weekend makes clear we are talking about "computer-generated" material. Would you like to me to prepend a warning? What would that warning say?

Re: London workshop on computer-generated session music

"I am sure not everyone here at the session.org maintains such an absolute position. The professional musicians with whom we are working are quite open to these ideas…."

With all due respect, Bob - Having professional musicians like "master Irish musician (your words) Daran Banarse" and Ensemble xy being open to the ideas is not impressive. Banarse apparently does some work on the "Irish Melodeon" and Ensemble xy is, what, some kind of classical/experimental troupe?

I would like you to put your own name on these things, since you already did that for the one you’re calling "Chicken" in your promo writeup. Why not put your name on them all? Too modest? Put the team’s names on them. Too many? Maybe that’s the real issue, then. It took XX people to build this thing?

The other thing I’d like you to do is explain how this is going to contribute to ITM, and why you came here.

I can see how it might contribute to AI, or to your own resume.

The reason for the project, according to your promo?

______________

"We aren’t so interested in whether a computer can compose a piece of music as well as a human, but instead how we composers and musicians can use artificial intelligence to explore creative domains we hadn’t thought of before. This follows on from recent sensational stories of artificial intelligence making both remarkable achievements — a computer beating humans at Jeopardy!"

_____________

That sure sounds like the project is intended to further the use of AI in creative domains. Fine - something new and cool to do with AI. But I’d simply like you to leave ITM out of it.

Posted by .

Re: London workshop on computer-generated session music

… so with that, I’ve rattled on long enough. I think I’ll hop off and play some tunes.

Posted by .

Re: London workshop on computer-generated session music

@Ergo
Are you convinced that, by either by definition or otherwise, AI is incapable of contributing to ITM as played in session, else scared that it might? What if in future, compositions could rival those of Seán Ryan [/insert favourite composer here]? Would you stop playing your one of your favourite tunes if you discovered it was composed by AI, else not wish to play an AI tune which you [otherwise] find most compelling?

Re: London workshop on computer-generated session music

I’m in Penrose’s camp when it comes to AI. The machines will crunch the numbers and do the big complex stuff that they do well. At the end of the day, though that may be pretty damn smart and several lifetime’s graft for a monkey, the machine will only every imitate art as it will only every imitate what we understand as consciousness and creativity. A machine will never have emotions, empathy and the need to express itself or unburden. The stuff that art is made of. Be off computer man, the day an tablet does more than pop up a few dots at a session will be a sorry one indeed. Take your AI mimicry and shove it.

Re: London workshop on computer-generated session music

A fiddle is like a machine. Doesn’t mean it cannot be part of how a player expresses her or his art.

Posted by .

Re: London workshop on computer-generated session music

A monkey with a rock got lucky and instead of smashing another’s head made fire instead. A rock is just a rock.

Re: London workshop on computer-generated session music

What you’re doing Bob is vaguely interesting but not very meaningful. The tradition is much more than the sequence of notes, to me at least. Best of luck in your endeavours though. It seems meaningful to you so more power to you.

Re: London workshop on computer-generated session music

I’m struggling with this. Why would a computer want to compose music? What would be it’s motivation? Because unless it has some emotion it wants to express through the medium of music it won’t ever be making music, just stringing notes together.

Re: London workshop on computer-generated session music

> A machine will never have emotions, empathy and the need to express itself or unburden. The stuff that art is made of.

Yes, and by trying to do it with a machine nonetheless you discover where, exactly, the art begins and the algorithm ends.

> it won’t ever be making music, just stringing notes together.

So what? There are plenty humans who never achieve anything more either. Electricity is cheap.

Posted by .

Re: London workshop on computer-generated session music

Is it possible to limit the input of tunes? Instead of using thousands of abcs, from several transcribers, downloaded from readily accessible collections could the source be more controlled? Say about
1,000 abcs prepared by one transcriber. Some method attempting to use more (relatively) consistent data
than a model based on highly variable collections. How might something like that change the output?

I’m thinking of the saying, "Too many cooks spoil the broth."

Posted by .

Re: London workshop on computer-generated session music

Tunes are composed once. A lucky few may then be iteratively performed, heard, selected, learned/taught and indeed modified until either absorbed into the tradition, else lost into obscurity. The act of composition itself is almost outside the tradition for the majority, else forms a miniscule part of it. Who cares where the tune came from? Many ‘modern’ tunes whose composers are either living or relatively recently deceased are incorrectly attributed and/or titled.

Re: London workshop on computer-generated session music

Of course, preserving a tunes’ provenance is also part of the tradition.

Re: London workshop on computer-generated session music

Thanks everyone.

Ergo: "How is this going to contribute to ITM?" That is a very good question, and it deserves a long hard think. Anyhow, I think your personal attack on the musicians with whom we are working is unfortunate. Look at our workshop page to learn more about who these musicians are: http://www.insideoutfestival.org.uk/2017/events/folk-music-composed-by-a-computer. At the very least, we are contributing to ITM by hiring these musicians to play ITM — not only at the workshop, but also the May concert. The workshop is also fully booked! (Will you be there?)

Steve T: Thanks for your perspective and friendly "shove". 🙂

joe fidkid: I can’t agree more that "the tradition is much more than the sequence of notes". In our work we try to stay consistent in describing our system as a transcription generator, not a music composer. For a more thorough discussion of why we are doing what we are doing, see this comment thread: https://thesession.org/discussions/37800

Mark M: Can’t music be about things other than emotion?

Calum: I really like your description, "by trying to do it with a machine nonetheless you discover where, exactly, the art begins and the algorithm ends."

AB: Indeed. For instance, now that the system understands rudimentary principles of ABC, we can fine tune its training on Dow’s 60. We have tested this with dodecaphonic music: https://highnoongmt.wordpress.com/2016/12/24/taking-a-christmas-carol-toward-the-dodecaphonic-by-derp-learning/

Re: London workshop on computer-generated session music

I think it would be interesting to find out whether "Bob’s Tunes" are really that much different from the tunes in the database on thesession.org . Highlighting specific tunes in a single thread doesn’t invite impartial feedback.

Bob, just for the sake of experiment, if you have the time, patience and resources (and assuming Jeremy had no objections), you could take a random batch of tunes from here, and an equal number of your generated tunes. Strip out the titles, and present them in random order, as quantised MIDI files with a decent piano sound. That way people would be forced to listen without prejudice, and so might be able to give more impartial feedback (and possibly spot a common element in either ‘type’).

Of course, there will always be the odd smart-arse who knows almost ‘every tune in the book’, but still … 🙂

Just an idea.

Re: London workshop on computer-generated session music

Bob it’s good to see you have a sense of humor, another thing machines have yet to master, despite the joke generators.

" Of course, there will always be the odd smart-arse who knows almost ‘every tune in the book "
Now that’s a joke a computer would have struggled with.

Re: London workshop on computer-generated session music

>> "Mark M: Can’t music be about things other than emotion?"

I don’t think it can, no. What do you think music could be about, other than emotion?

I can write a tune that conveys a feeling of love, but I can’t write one that says ‘Will you go to bed with me?’ To do that I would need to add lyrics. Like all forms of art, music is a channel of communication between artist and audience, and in the case of instrumental music that communication is limited to vague emotion - the sorrow of a lament, the confidence of a march, the reckless joy of jigs and reels. Take away the composer and there is no communication, the music becomes meaningless. If an artist splodges paint on the floor it is art, and he will happily tell you why he did it, what it means, and what he is trying to convey to you. If a tin of paint falls over and splodges paint on the floor it’s just a splodge of paint.

Instead of instrumental music, program your computer to do the same thing with the words of songs: take random words and phrases from a database of songs and use an algorithm to re-assemble them. The result will be speech, but it will be totally meaningless. It is the same with instrumental music: unless it is conveying something that the composer and performer want you to experience it is totally meaningless and pointless.

Re: London workshop on computer-generated session music

Just listened to The Drunken Pint clip - I’m afraid I’m with someone up above; it just sounds like what it is: a bunch of random notes and phrases cobbled together to no discernible purpose. Now, I’ve heard human-composed tunes that have the same effect, but at least I’ve been able to assume that, behind the failed tune, there was some idea or feeling that the composer was attempting to get across but that could not breach my tin ear and thick skull. Or that, sometimes, what to me was failed tune, was composed to fulfill the needs of some certain dance, game, chore, etc., which it perhaps did admirably.

Re: London workshop on computer-generated session music

Count me among the unbelievers. I listened to "The Drunken Pint" and it’s not even a tune. The second tune, the jig, the first part works, although it is very pedestrian, but the B part makes no sense. Didn’t bother with the 3rd "tune."

Re: London workshop on computer-generated session music

Ergo: "How is this going to contribute to ITM?" That is a very good question, and it deserves a long hard think. Anyhow, I think your personal attack on the musicians with whom we are working is unfortunate. Look at our workshop page to learn more about who these musicians are…."

Bob, you didn’t answer that very good question, nor will you. You’ve put ITM to use for your AI project, so don’t pretend this is intended to benefit ITM.

I attacked your musicians? You are the one who pointed out that "The professional musicians with whom we are working are quite open to these ideas…." and so were using their professionalism, or something, to bolster your case. I went to the workshop page to see who they are, and I went to their websites. Then I pointed out that they have nothing to do with ITM.

Your man Banarse can play the melodeon - I found a clip of him (below), but he also records the "hum" of the Tate Modern art gallery, writes music for TV, teaches Alexander Technique, and does a bunch of other things. And in this video he plays what he calls a "traditional style Irish reel" that he composed, rather than just playing a trad tune. He’d clearly be open to playing in your concert.

https://www.youtube.com/watch?v=eDs2_pirUmM


I’m sure your musicians are very good at what they do, but they have no skin in the ITM game.

The issue I have is still the same. You’ve put ITM to use for your own purposes, so my question remains: how does this knowledge, this modeling, benefit ITM?

You said above, "That is a very good question, and it deserves a long hard think."

Let me know when you have the answer.

Posted by .

Re: London workshop on computer-generated session music

Ergo: You posed a good and necessary question, but I really can’t believe how vicious your message is! I understand that you are passionate about music, but why do you feel you must tear me down and berate people who make a living by playing music, i.e., "professional musicians." It’s vile, really. Time to leave it there.

As to your good question, which other’s surely have, I have already made many suggestions in a previous thread (https://thesession.org/discussions/39604), e.g., automatic transcription systems, tools for composition, music education, … who knows! That’s the nature of basic research. The ABC here at the session, and from other resources (there’s a huge Swedish folk music ABC compilation project in the works), serves as a starting point for many other things. Google’s magenta project (https://github.com/tensorflow/magenta) is doing the same, but with thousands of MIDI files from a variety of sources.

Jim Dorans: Greetings again! Good idea on the experiment. We are preparing something like it, but avoiding MIDI completely. There’s so much more to this music that is not in the dots. Stay tuned! 🙂

Steve T: Thanks Steve. I’m a fan of Penrose too. 🙂 Not a believer in "the singularity" at all. More afraid of human stupidity than artificial intelligence. 🙂

meself and 5stringfool: Thanks for listening! A friend of mine said his wife overheard The Drunken Pint and said something like, "It sounds like a tune but it doesn’t!" I love that description. For me it is like a few loose pieces of meat in between fat and gristle. Doesn’t really hold together, like a drunk mess. 🙂 Yes, The Glas Herry Comment part A is rather pedestrian. Good description! What about the B part "made no sense"? One thing I see is the relationship to the A part is limited to just the tonality, meter, and a bit of the rhythm. It begins in the dominant, moves to the five of five, and then resolves in the tonic, which is nice compared to the rather static harmony in part A.

Re: London workshop on computer-generated session music

Bob, I can understand why musicians of any style are outspoken when it comes to strangers suggesting a relatively new approach, implying he or she may be contributing something to the style and potentially exploiting the music. There is legitimate cause for concern and skepticism. Though ultimately I think it comes down to either rejecting the strangers contributions or becoming familiar with her or his’ intent.

I’m assuming your intent, in the near future, is to work with more players and composers to see how they might, or might not, want to play just a few of the thousands of computer-generated compositions. For myself it would be finding those few compositions which may become tunes I might play.

With that in mind I went to your collection, moved the PDF beyond the 2,000th tune (to #2056), looked at that one, and scrolled backwards until I came to one which I thought I might play. It’s #2047. While it might not ever get played in a regular session or performance I would play it with other musicians in a session were it introduced. Point is this is a potential starting point for me. While I won’t play most of the computer-generated compositions in the collection I can see some potential, however small, in the lot.

Posted by .

Re: London workshop on computer-generated session music

[*Jim Dorans: Greetings again! Good idea on the experiment. We are preparing something like it, but avoiding MIDI completely. There’s so much more to this music that is not in the dots. Stay tuned! 🙂*]

Hi Bob! Yes, I agree, in principle - but good quality quantised (humanised) MIDI with a good piano sound is one of the best ways of demo-ing large amounts of music with a very ‘neutral’ sound (at least for the experiment I proposed). That rationale applies to all types of tunes, not just ITM.

If you can ignore all the current negativity going on here, just for the moment, I have a suggestion for you. I picked the "Glas Henry Comment", printed it off and played through it. The first thing that struck me was that some of the notes simply do not fit - their placement makes for obtuse-sounding phrases. I marked my recommended changes and copied it here :

http://worldfiddlemusic.com/guest/glas.jpg

At the time of your last thread, I scanned through a stack of your tunes, and this seemed to be a common element. None of the ones I looked at were "just a jumble of notes", though. If you actually sit down, read the music and play them through, what does pop out is what I said earlier - badly formed phrases, or just "bum notes" that make the bar(s) incongruous with the rest of the tune.

If you would play that tune through on your chosen instrument, first in its original form, then the amended version, then I think you would hear a big difference.

Re: London workshop on computer-generated session music

>>"There’s so much more to this music that is not in the dots. Stay tuned! 🙂"

I think you might be losing your objectivity here. How can there possibly be ‘more to this music that is not in the dots’ when the dots are generated by a machine? Are you suggesting that the machine has intentions as to how the music should be played that don’t conform to what it has written?

Given that you are using human musicians not MIDI synths, I think what you are seeing is not how AI can produce usable tunes, but how skilled human musicians can take any old crock of poo and make it sound more or less musical.

Re: London workshop on computer-generated session music

Mark M, don’t you think Jim Doran’s concept was a way of approaching the collections objectively for comparison & contrast? I’m asking because you suggested Bob is not being objective.

Posted by .

Re: London workshop on computer-generated session music

Yes, I do think Jim’s concept would give an objective view. However, Bob’s concept of not using MIDI means that he would be judging how human musicians can interpret computer generated note sequences and give them musicality, rather than judging what the computer generates.

Re: London workshop on computer-generated session music

I think interpretation is the next, obvious step in appreciating new compositions.
Isn’t it?

Posted by .

Re: London workshop on computer-generated session music

Yes, but it is another step, and it’s not what is being evaluated here. A derelict building is a derelict building, it doesn’t become less derelict if Banksy paints a mural on it.

Re: London workshop on computer-generated session music

The Glass Herry original which Jim changed could in fact have been a "real" tune, but an incorrectly transcribed one - or for that matter, a composition with an odd note here and there.

Re: London workshop on computer-generated session music

I flicked through a few of the tunes in the PDF linked to above. The tenth tune was the first that I came across that felt like a proper tune, with some kind of fluency and logic to it, where the B-part actually felt like a B-part, where the notes seemed like they were where they should be. It was a reel, and it felt like it had been written as a reel.

That said, a very ordinary reel though, and one I wouldn’t otherwise have given a second thought to.

All the rest that I looked at just seemed too random - they weren’t simple enough, they didn’t repeat or echo phrases enough, they didn’t have a logic to their phrases and cadences. What surprised me was that they didn’t seem "of their type" enough (i.e. a jig or a reel or a polka etc). I would have thought that an artificial intelligence would be able to make formulaic "jig-like jig" without too much trouble. It appears not.

I only looked at ten though. Given that a machine can, of course, generate literally billions of tunes it stands to reason that a proportion of those billions of tunes would be jigs that feel like proper jigs, reels that feel like proper reels - simply by sheer weight of numbers.

Re: London workshop on computer-generated session music

Ah, the 1000 monkeys= Shakespeare theory.

Re: London workshop on computer-generated session music

If a said composer doesn’t have the gaul to post himself playing the said composition, then I can’t see why anyone would take them seriously.

So, saying things like "lets see what someone can do with this tune" or whatever is kinda condescending.

The idea of AI playing tunes is great. I love the idea. The idea of any composer, AI or not, telling me how the tune goes or composing when they don’t know themselves is disturbing and would spark off a chaos of cacaphoneous clatter. my two bits

Re: London workshop on computer-generated session music

Thanks all!

Jim: Great suggestions on ways to adjust The Glas Herry Comment! It will take me some time to correct my fingers, but I will post a video comparison soon. 🙂

AB: Indeed, we intend to pursue this line of research and dialogue with practitioners in a concentrated manner. A grant proposal is in the writing! Unfortunately, a lot of research in applied machine learning takes data from somewhere without regard for its collection, uses it to show the "success" of a new method, and then moves on without asking people actually working in the domain of that data about the practical significance of what they have done. That requires a lot of hard work, and is a reason why I bother to post to thesession.org, organise workshops and concerts, attend some local sessions, etc.

Thanks for having a look through the first volume of folk-rnn tunes! #2047 is a really odd one. Measure 6 is rotten to me. The first and second "endings" are not endings at all. The system seems to just be rambling, but does form two 8-bar repeated sections, with some repetition and variation in the first part. Again, we are more interested in this generated material as points of departure, rather than destinations. 🙂

Mark M: We try to keep a clear distinction between transcriptions found here at thesession.org, and the _music_. When I wrote, "there is more to this music that is not in the dots" I was referring to traditional music, not to our system’s invented pieces. 🙂 Our system doesn’t compose music: it generates ABC.

AB, Mark M. and Jim: You have identified a point where we feel damned if we do and damned if we don’t. As I hint to above, ABC is just a bare-bones description of what to play. A hornpipe may be notated straight, but it certainly isn’t played that way. Missing too are accents, phrasing, dynamics, ornamentation, combinations of tunes into sets … everything that is really interesting, and totally out of reach of any computer system. So, if we compare MIDI-synthesised ABC, are we then comparing "music" because it is now something we can hear? On the other hand, if we have ABC interpreted by a musician trained in the conventions of session music, and then compare those recordings, are we then really comparing the algorithms independent of the musician? (In either case, what is the question you think a test participant should answer? Identify the computer generated tune? Identify the tune that best captures folk conventions?)

Nonetheless, we are not interested in whether our system can generate ABC that can be MIDI-synthesised and mistaken for "real" folk music. We don’t care whether our system has composed or can compose the best trad tune in the world. We have many more interesting questions to ask, and just as many ways to evaluate our system, for instance, see my recent blog post: https://highnoongmt.wordpress.com/2017/03/19/benchmarking-music-generation-systems/

Mark M: Really like your "derelict building" metaphor. We are certainly not advocating anyone take up residence in these shoddy creations of our system! 🙂

Matt Milton: Thanks for having a look! It can be hit and miss throughout. Your point about the ability to create a formulaic jig is good, but don’t throw the baby out with the bath water! Our system is built from one kind of statistical machine learning, which begins from the ground up. The system has not been taught anything other than to reproduce the next symbol given the previous ones. We impose no rules from above, e.g., stay to these pitches when in this key, trigger a measure bar every 6 beats, or create two repeated 8-bar sections and then end the transcription. Even so, it is neat to see that the system gives the appearance of being able to count, repeat and vary material, create an AABB structure, and produce some basic cadences.

Re: London workshop on computer-generated session music

Bob Strum, what would it take, hypothetically, to train your system from a small set of jig transcriptions (say 20-50 tunes) and begin generating computer compositions?
I’m thinking transcriptions by one transcriber, or checked by one transcriber. John Chambers has an index of transcriptions from O’Neill’s 1850.
http://trillian.mit.edu/~jc/music/book/oneills/1850/F/0701-0720_lw.abc
http://trillian.mit.edu/~jc/music/book/oneills/1850/F/

Posted by .

Re: London workshop on computer-generated session music

…or an even more limited sample: a number of jigs in the same key, with perfect AABB structure (for simplicity - without first and second endings, no pick up notes before the first bar, just like the Glass Herry Comment).

Re: London workshop on computer-generated session music

Your research makes me neither inspired nor does it cause me anxiety in answer to your other blog.

In an attempt to understand, I’ve read your most recent abstract. My questions:
1) What is the initial "kernal" for this folk tune Matrix that you’ve devised?
2)Are you using the assumed "right eigenvector" model or is it bifurcated?
3) Why would your research concern itself with human anxiety or inspiration?

All the best of luck

Re: London workshop on computer-generated session music

"Ergo: You posed a good and necessary question, but I really can’t believe how vicious your message is! I understand that you are passionate about music, but why do you feel you must tear me down and berate people who make a living by playing music, i.e., "professional musicians." It’s vile, really. Time to leave it there…"

Bob: if you think I’m being vicious and vile and I’m berating and tearing down people… you’ve lived a sheltered life. I’ll leave it there.

Posted by .

Re: London workshop on computer-generated session music

The more I think about this, the more I think that employing a mathematical algorithm to generate tunes isn’t all that different in principle to methods sometimes used in the past to create traditional music. There is precedent for using both random and structured processes as a jumping off point for tune creation.

A tune can be inspired by occurrences in the natural world:

chaotic baying of hounds on the hunt
the call of a morning lark
rolling waves
a whistling tea kettle

What you’ve got isn’t as random as throwing a handful of coins in the air and notating the jingle as they hit the floor. It’s closer to the sound of a breeze through wind chimes made of tubes tuned to a scale.

What seems to have been generated so far still feels pretty random. Scattershot that relies on sheer volume for successful hits. The 100 monkeys, 100 typewriters, 100 years scenario. I’m intrigued by the idea of limiting the scope. Could you construct rules that would yield a simulacrum of a Paddy Fahey or Ed Reavy tune?

Re: London workshop on computer-generated session music

"Could you construct rules that would yield a simulacrum of a Paddy Fahey or Ed Reavy tune?"

Can we not technologically tamper with everything that is good and pure in this world?

A computer farting out generated tunes in some academic lab somewhere is the beginning of the end. I’m amazed people are even giving this guy’s work serious consideration. The sooner this experiment is confined to an anonymous university archive the better. This Sturm guy would be well advised to spend his time practising his large accordion and then he may have an idea what it takes to come up with a traditional
sounding tune and then play it with some degree of style

Posted .

Re: London workshop on computer-generated session music

The sky is falling!

Re: London workshop on computer-generated session music

whats up with all these big accordion playing genius programmers who like to noodle with the bits?

Re: London workshop on computer-generated session music

I appreciate everyone’s input. For what it’s worth I am trying to create a dialogue with the original poster, Bob Sturm. Bob, thank you for responding to members’ questions, mine included. My interest in the OP is for the sake of discussion, which I hope to do with an open mind. I have my biases and am willing to bring them into the discussion; if that is essential. My intention is to progress toward a two-way dialogue with honesty and curiosity.

Ben

ps ~ Sorry for misspelling your last name earlier, Bob.

Posted by .

Re: London workshop on computer-generated session music

Thanks for even more useful input everyone!

Eddie mcguiness: Thanks for having a look! I’m afraid I don’t understand your first two questions. I’ve never described our system using the words "kernal", "folk tune Matrix", "eigenvector" or "bifurcated". We are using long short-term memory networks, which involve a sequence of linear and non-linear transformations. They project a sequence of 132-dimensional coordinate vectors into spaces of 512 dimensions, and then eventually back to a 132-dimensional space modelling a conditional probability distribution. When you say "most recent abstract", do you mean our article from last year’s Conference on Computer Simulation of Musical Creativity (https://csmc2016.wordpress.com/)? With regards to your third question, our work is situated in the research domain of computational creativity. This naturally leads to questions of the nature of creativity, and what it means when machines appear to do things that one would say are creative if a human were to do them. This field is not limited to just art, for instance proving mathematical relationships involves creativity. So, it’s not that our research is concerned with human anxiety or inspiration, but we wonder if those could be dimensions along which one can evaluate creative systems.

joe fidkid: Thanks for the picturesque and sonic metaphors! Statistical machine learning attempts to address problems that are difficult to address with a set of clearly specified rules and exceptions. The rub is that one must have a lot of data expressed in a suitably rich way, a model that is complex enough to embody the latent rules, and a training method that is able to discover those latent rules. "Deep learning" attempts to combine the expression and modeling into one.

Along with AB’s suggestion on focusing the training of the system on a specific subset, it is completely possible. We have done this recently, moving the system trained on thesession.org transcriptions to transcriptions of basic tone row compositions. It is really funny to hear it drop of the cliff of tonality (https://highnoongmt.wordpress.com/2016/12/24/taking-a-christmas-carol-toward-the-dodecaphonic-by-derp-learning/ ). What is interesting is that the system learned the concept of tone rows, but it would just create a sequence of four different tone rows. It didn’t figure out transformations of tone rows, e.g., inversion and retrograde.

In a different direction, we want to move the system toward the "voice" of a particular composer. For instance, my collaborator Oded is a professional composer, but has only a limited collection of works. Our system has learned some rudiments of ABC from thesession.org transcriptions. So, we are starting from its current state, and adjusting its weights using transcriptions of the smaller subset of Oded’s music.

This use of "leverage" appears in state of the art speaker recognition methods. We train speech models using massive datasets of speech sounds to get an idea of how speech is distributed in some feature space. Then for a specific user, we take a few recordings of their speech, and create reasonable detection functions that encompass the aspects of that person’s voice from the "universal background".

irishfiddleCT: Thanks for having a listen! Your question reminds me of when my dear mother once asked me, "Why do you have to analyse everything?" I can understand the fear that analysis will dispel magic or make something less valuable. In all my life of "over-analysing" things though, I have yet to see the fear come true. Appreciation grows with analysis.

AB (Ben): Thanks for providing good feedback and avoiding ad hominem and poison-the-well attacks. As I mentioned before, I am posting here because you all know better than me, play better than me, and embody a great resource of knowledge of a living tradition! (No worries about the name! 🙂

Re: London workshop on computer-generated session music

What would happen if you told you AI which of the "one in five tunes are actually fairly good" ?

I wonder what responses you would get if you posted some of the ‘top picks’ here as ABC and lied "this tune came into my head please can you tell me if I composed it or am partially remembering something else?"

Re: London workshop on computer-generated session music

Thanks David50.

We have been thinking about how to include such a "critic" in the training of the system. It’s not at all obvious how to do that!

That’s an interesting suggestion you make, but it is a bit too deceptive! 🙂

Re: London workshop on computer-generated session music

Great music! Apart from enjoying the computer-generated tunes, the clips clearly show good session dynamics.

Re: London workshop on computer-generated session music

Here’s my initial impression from Set #1. Could you try a set with more melody players and lose the guitar? No offense to the accompanist, but more melody players wouldn’t hurt your session. It might help.

Posted by .