Training on 12bits audio instead of 8bit? (Question, what do I need to change?)

[artificalaudio]

Hi,

I really really like the Sashimi, it's raw paper. Thank you.

I'm going to replicate to get my head around it, I'm thinking it should work with an expanded vocabulary and 12bit audio 4096 values instead of 256 for 8bit.

If I replicate the Beethoven experiment could I simply change the quant depth to 12 in the beethoven.yaml file?

Or is there an additional specific vocab setting I would need to set/change? For instance usual transformers a vocab size is set/known in advance.

Or is it just the quant variable in the beethooven file that needs changing and it'll just sort itself out?