Don't update duration if last timestamp is timestamp_begin

[vickianand]

There is bug causing start and end timestamps of the segments to be same when they shouldn't be.
If the only timestamp-token found is the tokenizer.timestamp_begin then we shouldn't update the duration (to 0). If there is really no speech, then anyway the segment won't be appended by the add_segment function.
Before this fix:

After this fix:

[jongwook]

Thanks. I think it's another failure case where Whisper didn't sample any intermediate timestamp tokens. Would you be able to upload/email the audio file?

[vickianand]

Here is the audio file: https://drive.google.com/file/d/1WG17h-3U565uWpl_jW2bkO3eS_l73H7f/view?usp=sharing
For the first three 30-sec chunks, it doesn't predict intermediate timestamps when run with:
--model medium.en --language en --initial_prompt "um, uh"