This week I wrote my first software review for MusicRadar, assessing a web app that promises to create sheet music from audio recordings. TL;DR it doesn’t do a great job.
One of my grad students is earnestly convinced that in the next five years, this problem will be cracked properly, that you will be able to drop an mp3 into an app and get a well-formatted lead sheet. I think it’ll get from 30% accurate to 75% accurate and then top out. The formatting is going to be harder than the musical content. That’s the hard part for humans! Deciding what information to include and what to exclude and where to place it on the page is an art, not a science, and AI sucks at art.
I gave my transcription students an assignment to make a one page lead sheet of “Space Oddity” by David Bowie. They didn’t have to figure out the song, they could use whatever references they wanted, they just had to fit all the information that a session musician would need on one page. They struggled! The good solutions were all quite different from each other, but some kids just couldn’t make it work at all, or sacrificed legibility for space efficiency. Hardly any of these kids have trouble transcribing, it’s the organization and presentation that’s hard. But that’s important!
I get people saying, oh, who cares, bad AI output is close enough. But reading a bad chart or lead sheet is much harder than reading a good one, so only advanced musicians can do it, and these are the people who least need AI to begin with. I don’t foresee any demand for poorly formatted and inaccurate music notation.
Songscription’s failures are quite interesting! The computer detects pitches and timing extremely accurately but it has trouble interpreting that information. Even in simple music, it does a bad job determining time signatures and metrical placement of notes. I got all this output where all the pitches and durations were accurate but everything was be an eighth note late. This is so close to being perfect and yet is totally worthless for actual use. I mean, sure, a pro could export that chart as music XML, bring it into an editor and fix it, but, like, a person on that level doesn’t need AI transcription to begin with.
Automated pitch and timing identification is quite useful for advanced applications! If these people weren’t promising readable sheet music I’d be very impressed by this product. Because who is this app for? “I have a gig tomorrow, I don’t know this tune, let me have the AI spit me out a chart.” If the chart is a mess, then the amount of work it would take to fix it quickly exceeds the amount of work it would take to just learn the song the old fashioned way.