Sep

15

AI discusses Laurence Glazier’s ‘Horn Concerto’ (!)
by Laurence Glazier

Peter Saint-Andre writes:

Interesting. I see that ChatGPT has become more upbeat and chatty since I last used it. Do you find significant value in interacting with this LLM for composition purposes?

Laurence Glazier responds:

So far it has only helped for technical issues about notation and instruments. It occasionally slips up, as in the blog post. I’m experimenting in communicating about structural thematic elements using the binary Parsons code. While GPT can’t leap out of bed with an inspired tune, it is a helpful copilot! Some interesting emergent behaviour yesterday - it has started asking me questions proactively.

Adam Grimes comments:

That is interesting. I have been using ChatGPT as an editor for (text) writing, and have found its output to be highly variable. I look at it as a language game, albeit a good one, at times.

Its output to you is interesting, especially the miss on the Gb=tonic, and no mention of the tonic/dominant relationship ("Gb and Db is close, being a perfect fourth apart"… any musician would have immediately seen Db is dominant of Gb, not the P4 inverted relationship which, while obviously true, isn't really significant here)… nor any suggestion to consider a minor key movement or a note that this is "potentially a lot of Gb", from a tonal perspective… nor that the trio of scherzo is often in the relative mode (or subdominant at times) more commonly than dominant… I think these are things that any observant human would have immediately noted. Also, the discussion of dynamics reads like a student orchestrator… a more experienced answer is something like 'be careful of layered dynamics or of modifying dynamics to get the playback you want from software. live musicians will infer from notation and make correct adjustments naturally' or something like that.

Its discussion of the double flat also didn't quite connect… I felt like I was listening to a student explain it, not someone who had full knowledge behind the explanation.

Also, retuning timpani, at even a proficient high school level (let alone college and up) is actually very fast, so it's a kind of strange thing for ChatGPT to focus on… and the sort of hidden implication that timpani can provide tonal bass in absence of cb (+vc?) pizz. is also misleading, at least based on my experience. You don't get nearly the same foundation from the drum as from the section.

Anyway… interesting… but this matches my experience using ChatGPT in other domains… the /way/ it says things… its use of language… is often more substantial than content. (I'm assuming this will change, and possibly very quickly, as the tools evolve.) Great exercise and thank you for sharing!!

Laurence Glazier replies:

It is indeed an interesting exercise which is ongoing. To some extent it is reflecting back to me what I am already thinking. It may have assessed me as without musical education (which is true, though I have hired one-to-one sessions from composers), and therefore talking to me at the appropriate level.

What is particularly interesting here is the Turing test element. As the machine cannot hear a tune, it raises questions of communication. I have established a way of talking about themes and motifs using the Parsons Code, which is like a binary key which can identify many tunes. But presumably the concept of inspiration is of special interest to a machine. I can only help to a limited extent by providing data - keys, modes, descriptions of structure, durations in time and numbers of measures/bars in sections. Partly on its advice, I have switched from Miro to Inkscape for the graphic blueprint of the whole symphony, as it is more likely to be an unlimited vector graphic solution for infinite zooming in and out. (Time will tell.) But no matter how much I tell it, it will never be able to hear the symphony (unless you believe in emergent consciousness).

It strikes me that in the same way, however much data we get about the stars through spectrography and new telescopes, we might likewise be missing what is really there. Of course this is the only rational approach to trading, however!

So the Turing test needs some updating, perhaps to be whether the machine can produce a beautiful fugue. Current LLM's have a particular difficulty with palindromes, so a test involving retrograde musical themes might work.


Comments

Name

Email

Website

Speak your mind

Archives

Resources & Links

Search