Question about batch_bins when batch_type is numel

[D-Keqi]

Dear Team,

In the speech translation task, when the type of speech is raw, is batch_bins equal to the sum of the speech length (i.e. audio samples) in each batch + the length of the target-language text + the length of the source-language text?

I did some statistics and it doesn't seem to be the case. I want to know how much data is contained in each batch when I set batch_bins to for example 4000000.

P.S. I'm using ESPnet2.

Thanks a lot!