Anyone tested Google Magenta: "Onsets and Frames: Dual-Objective Piano Transcription"

• Apr 5, 2018 - 23:02

Thomas mentioned magenta in chat, so I searched and found this recent blogpost from Feb 2018:…

(searching musescore forums for magenta, I see this comment by lasconic but nothing else).

I might try this to see if it can run on my labtop with my piano.


So I tested it with a 30-minutes recording of piano improv, and it analyzed it very quickly, and the midi ons and offs seem very accurate to me. It doesn't generate sheet music, so unless you are playing to a click or have some other means to detect beats, the midi won't look nice when imported into musescore.

In reply to by Nicolas

That would be much better. I will probably try that later this week with a click, and then also try without a click but with a steady tempo as I can.

Even though the analyzed midi was almost exact, the problem was importing that midi into musescore of course generated garbage because of no knowledge of when the barlines and tempos were.

In reply to by ericfontainejazz

As a test comparison, I took the 30 minute improv and imported into Anthem Score...took a couple hours using all my system resources to generate the musicXML file...and that imported into musescore and I could upload it:

Of course the notation isn't that accurate because of the unsteady time and because of ornaments and what not which Anthem Score doesn't understand...but at least i got something...and the neat thing was anthem score did actually understand tempo changes to some extent (most measures it had a new tempo marking). And then I used the auto sync to youtube ability, and so I somehow have the mediocre anthem score notation somewhat synced to the original audio.

Do you still have an unanswered question? Please log in first to post your question.