Non-streaming Conformer model with pruned_rnnt_loss always emits the first non-blank characters on the very first frames.

[guoyifan97]

I trained two offline reworked conformer models on my own Chinese data using pruned_rnnt_loss and standard rnnt loss (warp-rnnt==0.7.0) respectively following pruned_transducer_stateless5. However, I still experience the issue where the first word timestamp is aways zero with the conformer + pruned_rnnt_loss. While the conformer + standard rnnt loss does not have this phenomenon.

I have seen the comments in :
#1347
#942
#923
k2-fsa/sherpa#52

But I still don't know how to avoid similar problems. Is there any way to solve such problems?
Thanks！

[guoyifan97]

I also found that using odinary rnnt loss + k2 simple loss will experience the same issue. But why?