ai4bharat

/

IndicF5

Like 372

Follow

AI4Bharat 2.27k

Hi AI4Bharat team, and thank you for releasing IndicF5 openly.

Before building on it commercially, I'd like to confirm one thing: was
IndicF5 pretrained from scratch on the 1417h corpus, or initialized from
the released F5-TTS checkpoints? Those are CC-BY-NC because of Emilia, so
I wanted to confirm rather than infer.

Two things suggest from-scratch, but I'd rather have it stated than assume:

• the checkpoint is model_1176000.pt — 1.18M steps is a pretraining-scale
  run, not a fine-tune
• vocab.txt shares F5's 2545 size but contains no CJK tokens at all,
  only Indic scripts and ASCII

A one-line confirmation would be very helpful, and I suspect others building
commercially have the same question.

Thanks again for the work.