Why use different hop size for the medium and large model?

[newoneincntk]

Thanks for sharing your amazing work.
I have a question: why are different hop sizes used for the medium and large models?
I try to train the large model with a hop size of 256, but the performance turned out really poor. Thank you.

[aask1357]

In my experiments, reducing hop size resulted in slightly better performance. However, the large model with hop=256 should not be poor. Can you share the configs (dataset, 16khz/48khz, etc) that you used?