Audio to sheet music - Recent News and a Ten Year Outlook

A few weeks ago, I read about a new AI model that can listen to music and read sheet music at the same time, instead of doing one and then the other. Last week, I read that MuseScore, free notation software used by millions of people, launched a tool that turns a recording into playable sheet music. On their own, those are just two news items. Put together with a couple of other stories from the same two weeks, they tell a bigger story: AI that turns a recording into notation is moving fast, from research lab to free app, and it's worth thinking about where that road leads over the next ten years, not just the next ten days.

Two separate things are becoming one thing

A piece of sheet music and a recording of someone playing it have always been two different ways of capturing the same performance. Turning a recording into notation, called transcription, takes a trained ear and a lot of patience, or an expensive piece of software (OMR) that isn’t 100% reliable yet. This month, two research papers showed that gap closing. One, from researchers in Singapore, won a best paper award for teaching a single AI model to understand written notation and recorded audio together, instead of using one tool to listen and a separate tool to write it down. Another paper introduced a big new practice set of popular songs (not just classical music, which is what most research in this area uses) paired with accurate sheet music, and used it to cut transcription mistakes by more than two-thirds compared to the best previous method.

That second detail matters for a simple reason: almost all the academic research in this field is built and tested on classical music, because that's where clean data already exists. But that's not what your students are recording on their phones in rehearsal. In a perfect world, every student audio recording would be clean and easy for anyone, including an AI, to evaluate. The real world gives us errors and noises in the background and restarts. The human element retains the advantage for real world application. 

A free tool that 44 million people already have open

Research doesn't change a classroom until it shows up somewhere students and teachers already are. On August 17, MuseScore, one of the most widely used notation programs in schools, launched a beta feature that turns an uploaded MP3, a YouTube link, or a recording into editable sheet music, for free. It also added a Shazam-style tool that listens to a song and finds a matching, ready-made score. MuseScore says it built and trained the underlying AI itself, and the company reaches about 44 million people a year. This isn't a small startup experiment. It's built into software many of your students and colleagues already have installed.

A competing company, Soundslice, kept improving its own version of this technology in the same two weeks: faster scanning of paper sheet music, better automatic labeling of instruments, and a small update that shows you exactly where the beat falls while you're lining up a recording with its score. None of those updates is a big headline by itself. Together, they show a real race happening in this space, and races usually mean faster progress and lower prices for the people using the tools.

Being honest: choir and ensemble music is still the hard part

I want to be straightforward here instead of just upbeat. Buried in that same Soundslice update was a fix for handling a part or voice that doesn't start until partway through a piece, something that happens constantly in choir music, where the vocal line often comes in several measures after the piano does. The fact that this needed its own fix tells you where these tools still struggle. A solo piano recording or a single guitar line is close to a solved problem. A full choir or ensemble score, with different voices entering at different times, is still where these tools trip up. That's exactly the kind of music many music teachers, including me, work with every day.

Proof this is more than a shortcut

I don't want the caution above to be the last word, because the same two weeks gave me a genuinely moving example of what this technology can do at its best. The European Commission highlighted a research project that used this same kind of AI to recover medieval Gregorian chant, music written down roughly a thousand years ago that hadn't been performable again until now. That goes well past a classroom time-saver: it's music history, brought back to life. When someone asks what this technology is actually good for, beyond saving a teacher some time, this a valid answer. 

Looking ten years ahead

Let’s play fortune teller: Over the next one to three years, turning a recording into notation stops being a beta feature and becomes something built into the free tools schools already use, and it gets noticeably better at handling popular music, not just classical pieces. 

Over the next three to seven years, as AI starts treating the recording and the written score as one single thing instead of two, the whole workflow changes: instead of recording something, then transcribing it, then cleaning it up as three separate steps, you'll work inside one document that's the sound and the sheet music at the same time. During that same stretch, expect the choir and ensemble gap to close as more real ensemble recordings become part of what these tools learn from.

By the ten-year mark, the question won't be whether AI can turn a recording into sheet music. It will be able to, reliably, for almost anything your students play or sing. What changes is how normal that becomes to use, more like a metronome or a tuner sitting on the stand than a piece of software anyone has to think their way through. Picture a rehearsal where a teacher pulls up the recording from five minutes ago right next to the original sheet music, side by side, so a student can see exactly where their entrance came in early or where a pitch drifted, not just hear that something was off. For students who learn best by seeing something rather than just hearing it, that side-by-side view becomes an extra set of eyes on their own performance, pointing straight at the measure that needs work instead of leaving them to guess from a comment alone. A trained ear still makes the final call, just with a visual partner standing next to it now. Getting programs ready for that shift now, instead of being caught off guard by it later, is exactly the kind of thing MusEdLab is trying to help music educators do, while there's still time to shape it ourselves instead of just reacting to whatever the software companies decide. 

Sources

1. NUS Computing, "NUS Computing Team Wins Best Paper Award at MMM 2026"
2. "Audio-to-Score Transcription using Pre-trained Features, Data Augmentation, and the New SheetSage-A2S Dataset", arXiv:2608.06165 (Aug. 6, 2026)
3. Muse Group, "Hear It, Play It: MuseScore Introduces New Smart Audio Recognition Features" (Aug. 17, 2026)
4. Soundslice, "Sheet-music scan improvements" (Aug. 13, 2026) and "New: See detected beats when editing syncpoints" (Aug. 28, 2026)
5. Repertorium project, team page and Salamanca Cathedral chant recovery; background via National Catholic Register. Originally surfaced via a European Commission LinkedIn post (Aug. 12, 2026).