Finetuning parakeet-tdt-0.6b-v3 - Guidance on data preparation

[Skaddd]

Hello Team,

First, thanks for the incredible work !

From this issue #14566 I understand that both canary-1b-v2 and parakeet-tdt-0.6b-v3 does not support TN as they were trained directly on unnormalised text.

So to finetune parakeet, I assume that the best way is to adapt my training data by applying ITN transformation. But would it sufficient to use pipeline from nemo-text_processing?

Or there are custom ITN rules that I will be missing ?

For punctuation, if my training data doesn't have it, I assume that the model will be biased during fine-tuning and incrementally output less punctuation tokens

Thanks in advance for your responses !