An option to FinalResult which would skip all the deallocations

[FredSRichardson]

This is really just a feature suggestion to add batch mode segment based decoding to the Recognizer.

Implementing this myself, I ran into the following problems:

• When you call FinalResults() to get the final results the call deallocates a bunch of things including the decoder which has to be reinitialized now between each segment adding unnecessary overhead.
• There is no way to call InitState() which means word offset times are adjust by however much data you have decoded up to that point.

I would suggest adding an option to FinalResult() which would skip all the deallocations and an option to Result() and FinalResult() which does not adjust start/end times to support batch mode processing. It would then be up to the user to periodically reallocate the decoder (for example to avoid memory leaks etc.).

[nshmyrev]

If you want to reallocate the decoder you simply reallocate the decoder. I don't see any issue. There is no overhead, decoder object is lightweight.

[FredSRichardson]

Yeah, that's not what I saw. In Python, I added a step deallocating the decoder and reallocating between each segment and this ran significantly slower.

[nshmyrev]

And if you just feed data after FinalResult? It should also work.

[FredSRichardson]

Yes, this works, bit you are taking an unecessary hit deallocating and reallocating the decoder. Also you have a time offset added to each word start time. It's just not ideal for segment based processing though it is great for streaming processing.

On Wed, Jun 22, 2022, 6:12 PM Nickolay V. Shmyrev @.***> wrote:

And if you just feed data after FinalResult? It should also work.

— Reply to this email directly, view it on GitHub https://github.com/alphacep/vosk-api/issues/1037#issuecomment-1163661266, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAPMHIMWZHF4N6KIQVKBCRLVQOFWNANCNFSM5ZRHEZHA . You are receiving this because you authored the thread.Message ID: @.***>