I think this is already possible with some manual effort.
However, AI companies likely avoid encouraging covers of copyrighted music to reduce legal risk. Record labels actively pursue unlicensed covers, and a company monetizing them would be an obvious target.
It would also likely make it easier to circumvent copyright detection algorithms.
That's why I decided to allow arbitrary order within the file. In this way you can group notes by instruments and the parser will deal with reordering them.
I also plan to create a flag for the CLI tool that reorders the lines within the mtxt file in such way, that notes are grouped by instruments.
ABC notation is more oriented towards traditional sheet music, with regular note lengths, standard Western tuning and a simple, readable syntax. It isn't meant for playing back music that sounds good to the ear. It's hard to catch the nuances of a real human performance with it, but it works well as a lead sheet for musicians. Its expressive marking are relatively limited and interpreted subjectively.
MTXT focuses on editable recordings of live performances, preserving all of those tiny irregularities that make the music human. It can represent arbitrary timings, subtle expressive variations and even arbitrary tuning systems. MTXT can also capture transitions like crescendos and accelerandos exactly as they happened.
My initial goal was to fix some mistakes in the MIDI files I recorded from my keyboard. I was also interested in making dynamic tempo and expression changes without dealing with complicated DAW GUIs.
Now I'm working on a synth that uses MTXT as its first-class recording format, and it's also pushing me to fine-tune a language model on it.
I considered it but decided against it in the first version, because specifying note durations is too tricky. It was more important to get the .mid -> MTXT conversion and live-performance recording working, where notes usually have irregular note lengths.
Representations like "C4 0.333 D4 0.333 E4 0.25" feel too hard to read.
However, AI companies likely avoid encouraging covers of copyrighted music to reduce legal risk. Record labels actively pursue unlicensed covers, and a company monetizing them would be an obvious target.
It would also likely make it easier to circumvent copyright detection algorithms.