Per social media posts I've seen, the new large model from OpenAI ("Astra") is capable of engraving scores in Lilypond based on image inputs. The most in-depth post I've found is here, from an earnest musician who says he or she is going to upload this to IMSLP once they have an account: https://x.com/aug5thmusic/status/2098469467973968206
I am raising this in the forum because I think it will be of general interest to IMSLP administrators and copyright reviewers. It is possible that a tide of machined-engraved scores is coming soon. If the cost of Astra is prohibitive now, it will not be in a year or two.
In light of this, the policymakers of IMSLP may wish to consider:
* whether having many AI-engraved scores is desirable (more scores!) or bad (potential quality issues)
* whether use of AI by users in engraving or for other purposes is a matter that needs a policy
* whether AI is good enough to explore using its assistance in copyright review; automation of some or many steps may now be possible
I have used AI on IMSLP myself — my use was for transcribing tables of contents (which I check, of course). My personal view is that there is a great deal of potential here, but things may change quickly enough that the people maintaining the site will need to start thinking about it now.
Expect AI-engraved scores soon
Moderator: kcleung
-
Fugalicious
- regular poster
- Posts: 22
- Joined: Wed Aug 29, 2018 4:55 pm
- notabot: 42
- notabot2: Human
Re: Expect AI-engraved scores soon
Hello! I'm Auggie, the editor of the Astra Edition of Weber Symphony No. 2. I'm happy to answer any questions you may have about my work. First, I want to assure you that I am an experienced engraver and this edition is as good as anything I could have produced on my own.
If IMSLP administrators want to institute a blanket ban on AI-assisted submissions, I understand. AI content has been quite poor for most of its history. But I can assure you that if I had submitted this score under my own name, without disclosing Astra's role, it would have been accepted in a flash. Any unbiased observer would see that it is a meticulously edited and engraved score.
For the past few days, my working habit has been to proofread the score every now and then, find things that need fixing, then submit the list to Astra for implementation. This is much easier than digging through the LilyPond code myself and collating every change across multiple parts and scores. During the initial stages, I relied on Astra to find and fix errors itself. Sometimes it submitted reports and I reviewed them. Other times I told it to go off and find things to fix on its own. But I always reviewed the results, and as the editor, I take full responsibility for every note on the page.
If you want to see the Astra Edition full score, there is a video on YouTube. I've made small corrections since uploading it, but this version is close enough to the final score to give you a sense of its quality:
https://www.youtube.com/watch?v=DOegZHxqBRY
If you want to see Astra in action, I recorded a video of my screen as Astra engraved a Haydn string quartet movement for me:
https://x.com/aug5thmusic/status/2098584544895025193
I include every prompt in the replies, along with a link to the GitHub repo (https://github.com/aug5thmusic/ai/tree/ ... a-haydn-sq) that contains Astra's engraving of the score and parts.
If IMSLP administrators want to institute a blanket ban on AI-assisted submissions, I understand. AI content has been quite poor for most of its history. But I can assure you that if I had submitted this score under my own name, without disclosing Astra's role, it would have been accepted in a flash. Any unbiased observer would see that it is a meticulously edited and engraved score.
For the past few days, my working habit has been to proofread the score every now and then, find things that need fixing, then submit the list to Astra for implementation. This is much easier than digging through the LilyPond code myself and collating every change across multiple parts and scores. During the initial stages, I relied on Astra to find and fix errors itself. Sometimes it submitted reports and I reviewed them. Other times I told it to go off and find things to fix on its own. But I always reviewed the results, and as the editor, I take full responsibility for every note on the page.
If you want to see the Astra Edition full score, there is a video on YouTube. I've made small corrections since uploading it, but this version is close enough to the final score to give you a sense of its quality:
https://www.youtube.com/watch?v=DOegZHxqBRY
If you want to see Astra in action, I recorded a video of my screen as Astra engraved a Haydn string quartet movement for me:
https://x.com/aug5thmusic/status/2098584544895025193
I include every prompt in the replies, along with a link to the GitHub repo (https://github.com/aug5thmusic/ai/tree/ ... a-haydn-sq) that contains Astra's engraving of the score and parts.
Re: Expect AI-engraved scores soon
That's just using ChatGPT for OMR, essentially. It's fine, as long as someone competent checks it thoroughly for accuracy and makes any necessary edits. There is no creative content whatsoever from the engraving program; it's not an AI-generated original work. It's doing OMR from score to LilyPond output.
I would caution that OMR is not really reliable by itself, but, if it's checked, then it's fine. (I would suggest that the LLM-written prefaces should be replaced, however.)
On the other hand, for totally separate reasons, no, we can't use AI for copyright review. This must be done by humans.
I would caution that OMR is not really reliable by itself, but, if it's checked, then it's fine. (I would suggest that the LLM-written prefaces should be replaced, however.)
On the other hand, for totally separate reasons, no, we can't use AI for copyright review. This must be done by humans.