fix: improve heuristic for not combining paragraphs#560
Conversation
rroohhh
left a comment
rroohhh
left a comment
There was a problem hiding this comment.
Can we get a test for this?
Also can you describe what you actually changed?
|
huh, it seems like this is actually also not yet working. A test document (https://media.ccc.de/v/39c3-denkangebot-rainer-mhlhoff-ber-ki-und-autoritre-sehnschte-im-silicon-valley#t=1811) now transcribes in a way that is not better and hangs the editor quite badly. Ill have to investigate some more |
anuejn
force-pushed
the
anuejn/fix_one_word_per_line
branch
2 times, most recently
from
97ecec2 to
2f5a402
Compare
|
should be good to go now :) |
rroohhh
left a comment
rroohhh
left a comment
There was a problem hiding this comment.
Do we remember, why we had the "until the end of the document" behavior previously?
|
The assumption was, that the whisper model will not get out of its corner of not producing sentence breaks anymore. This is mostly true but our solution lead to some even worse failure when the heuristic triggered accidentally. Basically we made a less-than-optimal trade-off. |
this changes the detection heuristic for overly long sentences from the previous "longer than 30s" to a word based approach of "more than 161 words". moreover it gets rid of the stateful approach that we would disable all recombination after the first overly long paragraph. overall this is not a perfect approach but produces saner results than before. this should temporarily fix the situation in which transcribee generates only one word per paragraph. the proper fix is resetting the whisper state from time to time and will land in the future.
anuejn
force-pushed
the
anuejn/fix_one_word_per_line
branch
from
2f5a402 to
a3eaa57
Compare
2 participants
No description provided.