cannot reproduce the paper results

[Sharpness1i]

Hi,there
when i tried to reproduce your paper results, i found that the segmentation module is barely able to effectively segment online input waveform which is very strange to me.
It often happens that the segment module "forgets" to segment even after finish reading a 3 second audio input, leading to the fact that the model acutally turns into an offline mode. The BLEU score and latency metrics cannot match with the model you have trained out
I strictly followed the training tips in READ.ME. Steps are rigorously implemented, like MT pretraining, offline training then finetuning using segment-speech option.
The dataset i adopted remains the same.
Would you kindly open source the checkpoint ? Thank you and wish your the best !

[krgy12138]

So am i.