Songscription launches an AI-powered ‘Shazam for sheet music’

A small company called Songscription launched last week with AI models that automate music transcription, turning an audio file of a song into sheet music within minutes. Operating on a freemium model, the product is geared toward both professional and hobbyist musicians.

“We hope to make playing music more enjoyable,” Andrew Carlins, CEO of Songscription and a student in Stanford’s MBA/MA in Education program, told TechCrunch. “We imagine a future where a rural Nebraska high school band teacher [will be] able to get sheet music for the songs their students want to play, [and] that said music will be arranged specifically for the instruments in the band and offered at the individual level of play of each student.”

At launch, Songscription can transcribe music for several different instruments, though the piano model is most reliable. In the future, the company hopes to add different transcription outputs (e.g., guitar tabs), as well as arrangements for a full band as opposed to just one instrument.

This kind of product could be useful for a musician who records a song they’re working on, then uploads it to get the sheet music — that way, they can skip the step of having to manually transcribe their work. And for those who can’t read and write sheet music, Songscription will also generate a piano roll, which shows a digital representation of the music being played on a virtual piano.

Users can also automate music transcription directly from YouTube links. Uploading a file requires users to check a box to confirm that they have the rights to transcribe the file, but it would be easy to simply check the box and get free sheet music for copyrighted songs.

“For music learners … since you are allowed to listen to a song, write down the notes by ear, and perform it on your home piano (as long as you don’t charge for a performance), it isn’t fully clear that using a tech-enabled platform to give you a head start crosses any legal boundary, although we understand the field is evolving and our application may enter a gray area,” Carlins said.

The legality around much of how we engage with these creative AI tools is up for debate, though recent court decisions seem to be favoring tech companies over artists. However, Songscription isn’t creating new, AI-generated music — it’s providing a tool for musicians to speed up the process of making their own guitar tabs or sheet music.

“Since our platform allows users to edit the scores, we position ourselves as an augmented music notation software that helps people speed up the process of transcription,” Carlins said.

The underlying architecture of Songscription’s AI model is based on a paper that co-founder Tim Beyer published, alongside researcher Angela Dai.

In order to get the training data necessary for creating this kind of AI model, Songscription works with some musicians who were willing to share or sell their piano performances and sheet music. The company also used public domain sheet music, though the majority of the training data is synthetic, Carlins said. In that case, Songscription would convert the sheet music into audio, then alter the files to simulate real-world conditions with background noise or reverb.

Just seven months after it was founded, Songscription has raised pre-seed money from Reach Capital and will participate in Stanford’s StartX accelerator.