Automatically refine word-level alignments from sentence-level alignments

[ryanfb]

First of all, thanks so much for all your work on this and making it open source! It would be cool if it were possible to do a fragment search using an existing SRT transcription without having to re-transcribe all of the audio in advance. One way to do this would be to use the existing sentence-level alignments to extract the audio ranges for sentences that match a search, then use vosk to transcribe just those audio ranges, then use the results of those transcriptions to extract the fragment-level audio.

[antiboredom]

That's an interesting idea - I'd definitely be open to experimenting with it... Alignment might also work here. alphacep/vosk-api#756

[cmprmsd]

It would also be beneficial to rely on the words sourced from the subtitle file. That way the detection quality could be improved a lot, right? I tried to implement @ryanfb's suggestion back in 2021 with pocketsphinx but the results weren't promising. 😢