Mimi structure and utilization

[RinRin-32]

Due diligence

• I have done my due diligence in trying to find the answer myself.

Topic

The paper

Question

To my understanding, Mimi encode audio into semantic-acoustic token via a single tokenizer. I'd like to ask if Mimi has a discrete number of token, what the size of the cookbook is and if it's possible to be utilized in TTS application. My main idea currently is to somehow train a transformer based model to generate semantic-acoustic tokens based on input text, speaker, speaker attribute, etc.