Training data used for the released checkpoint

[wincing2]

Thank you for releasing the code and the pretrained checkpoint. On what data was the checkpoint trained and how many hours are these data? Only LibriSpeech as stated in the paper?

[Aria-K-Alethia]

Hi,

Only LibriSpeech, of course. Everything stated in the paper is faithfully reflected in the code and ckpt.

[Liujingxiu23]

@Aria-K-Alethia I see you claim in the paper that "trained on 8 NVIDIA A100 GPUs" and "model converges in about 600k steps.". I want to know about how many days to train to achieve 600k steps ?

[Aria-K-Alethia]

@Aria-K-Alethia I see you claim in the paper that "trained on 8 NVIDIA A100 GPUs" and "model converges in about 600k steps.". I want to know about how many days to train to achieve 600k steps ?

I don't remember clearly but I think 2~3 days. If possible, I suggest you increase batch size to 16 or more for a faster convergence.