Documentation on MusicLang Tokenizer

[MikeMpapa]

Hi and thanks a lot for the great work! I was wondering if there is any detailed documentation on the tokenizer beyond what's in the technical blog? Also is there a way to add custom tokens based on metadata in order to control on inference?

[floriangardin]

Hi, thank you.
Adding custom tokens would imply retraining the model (because the vocab would change). Unfortunately, at this time I don't have a timeline for releasing the training code. You can check the tokenizer class to get the tokens structure, and check musiclang language documentation to see what the tokens means. Can you share what kind of tokens do you want to add in the model ?

[MikeMpapa]

I would be interested to control stuff like genre, mood, and similar tags of that nature that might be more abstract in nature but live in the metadata-land

[iplanwebsites]

Also looking to add control tokens so accessing the training code would be great.
Miditok is a good meta-library for miditokenizatoin. Would be nice to standardize your tokenizer in that lib.
As a proxy for harmony and general feel, I'd like to experiment providing note histogram as control-token and see how it can help shape the genre and harmonic blueprint of a song.