Re: Anybody else playing with GPT4 and Lilypond?

[Nick Bailey]

This is nothing to do with GPT4, or guitars, so might be considered off-topic :) , but I'm aware of this paper due to Percival et al., which addresses sight-reading exercise generation:

https://www.researchgate.net/publication/235925970_Generating_Targeted_Rhythmic_Exercises_for_Music_Students_with_Constraint_Satisfaction_Programming

On 30/03/2023 05:57, Mike Blackstock wrote:

re. "Anybody else playing with GPT4 and Lilypond?"

I'm very much interested in exploring its use to generate graded sight-reading material.

My own instrument is classical guitar and we're not the best sight-readers[1]... it would be

nice to have daily sight-reading exercises generated for practice, with midi. I could donate

the use of a QEMU/KVM server instance for working on a project of that sort.

[1] Guitarist John Williams:
"Another thing I’ve noticed in master classes, is that players will come on and play the most
difficult solo works from memory, and yet if you give them a part to play in one of the easier
Haydn String Quartets, as I often do, they’re lost in no time, and have a very poor sense of

ensemble or timing. Guitarists are among the worst sight-readers I’ve come across.
Julian Bream and I are both dead average sight-readers by orchestral standards,
but among guitarists, we are [considered] outstanding! "

https://guitarteacher.com.au/interview/john-williams-interview/

On Wed, 29 Mar 2023 at 18:44, Saul Tobin <saul.james.tobin@gmail.com> wrote:

I've seen some examples of other people succeeding in getting ChatGPT with GPT4 to compose simple music in other text based music formats. I've had limited success getting it to output Lilypond code. It is able to correctly structure the code with a score block, nested contexts, and appropriately named variables, and bar checks at the end of each measure. It seems to struggle to create rhythms that fit within the time signature beyond extremely simple cases. It also seems to struggle a lot to understand what octave pitches will be in when using relative mode.

It also seems to have a lot of trouble keeping track of the relationship between notes entered in different simultaneous expressions. Just asking it to repeat back which notes appear in each voice on each beat, GPT4 frequently gives stubbornly incorrect answers about the music it generated. This makes it very difficult to improve its output by giving feedback.

I'm curious whether anybody else has tried playing with this. I have to imagine that GPT4 has the potential to produce higher quality Lilypond output, given some of the other impressive things it can do. Perhaps it needs to be provided with a large volume of musical repertoire in Lilypond format.

--

https://blackstock.media