Why the OnlineASRProcessor.commited is always incremented?

[bxg]

Why the OnlineASRProcessor.commited is always incremented (it includes all commited transcripts from the beginning)?
As time passes, the "commited" grows large. Should we truncate it ?

[Dominik Macháček]

Hi,
thanks for feedback. This is an unresolved corner case. Yes, you can truncate it after this line

whisper_streaming/whisper_online.py

Line 269 in cc305b3

 self.commited.extend(o) 

, but there must remain 200 last characters. Or other size, I haven't experiment with the 200 here:

whisper_streaming/whisper_online.py

Line 245 in cc305b3

 while p and l < 200:  # 200 characters prompt size 

, it can be tuned/improved.

Does it the current version without truncating harm the performance? Do you have memory issues? Or is it just annoying in the log?
After how long audio/how many words? Usually nobody runs one audio processing so long so that's an issue.

[bxg]

Thanks. I just review the code and think it will be harmful to performance and memory. I think of its application in video conference.