Decoding won't stop if changed window size from 30s to less than 30s

[BuaaAlban]

I have tried to modify the window size from 30s to 10s, and change the related ctx length from 1500 to 500.
But I found the decoding with duration around or longer than 10s won't stop, which means no prediction of <|endoftext|>.
In the paper it says: "We break audio files into 30-second segments paired with the subset of the transcript that occurs within that time segment". I wonder whether there adds a <|endoftext|> token at the end of 30s transcript during training, and the encoder may tend to stop at end of 30 seconds.
Thanks for any suggestions or explanation.