Changing batch size to 16 or 32

[Peleg-Bruck]

Hi, I've read some of the issues here, and saw that you mentioned that you were suspecting that changing the batch size from 64 may cause the conversion to be suboptimal. Did I understand you correctly? My GPU cannot take a batch of 64 and I therefore changed it to 16. I trained for 150k iterations(starting from your pretrained 24khz checkpoint) and the conversions are not all that great. It sounds crisp, but the similarity is not very good. If you think the batch size is the culprit, should I keep batch size 64 and decrease the training wav size? Plus, given 10 hours of data, how many iterations do you think I should train? Thank you!