Undesired splitting with regroup

[darioai]

@jianfch , first of all, thanks so much for creating stable-ts! This thing is amazing!
Now, to my issue: I'm using regroup to keep segments at a max_chars = 65, which after multiple tries (I don't have any formal training in Python) is working as it should 👍
However, regroup is causing some undesired splitting to subtitles under the set max_chars, like the example below:

From:
00:00:00.096 --> 00:00:02.440
En un edificio como este es un juego

to
00:00:00.096 --> 00:00:00.196
En

00:00:00.768 --> 00:00:02.440
un edificio como este es un juego

Any idea on how to fix this?
Thanks!

[jianfch]

It seems to have split due to a gap from 00:00:00.196 to 00:00:00.768. Which methods did you use?

[darioai]

This is the command I'm using:
command = ["stable-ts", audio_path, "-o", transcription_path, "--model", "large-v3", "--vad", "true", "--segment_level", "true", "--word_level", "false", "--max_chars", str(max_chars_value)]
Is there a way to turn services off from the command?

[darioai]

I just added "--regroup", "sl" to the command and it took care of the problem. Thanks for the quick reply!

[jianfch]

When --max_chars or --max_words is specified, it calls split_by_length() after it regroups. Since --regroup true is the default, it will use the default regrouping algorithm which includes split_by_gap(). So disabling the default regrouping with --regroup false should achieve the same results with less compute. Note that --regroup "sl" will call split_by_length() without specifying max_chars or max_words which means it will do nothing.

[darioai]

Is there a way to disable split_by_gap w/o disabling regroup?
I

[jianfch]

Is there a way to disable split_by_gap w/o disabling regroup?

The default regrouping can be represent as "cm_sp=.* /。/?/？/,* /，_sg=.5_mg=.3+3_sp=.* /。/?/？", True, or a chain of regrouping methods in the section about regrouping. So simply remove sg=.5 to disable the split_by_gap(.5). The first .* /。/?/？ is also redundant so remove that as well. You end up with --refine "cm_sp=,* /，_mg=.3+3_sp=.* /。/?/？".

[darioai]

Thanks!!!!

[darioai]

Since different caption systems have different requirements for the max chars/line, I've been using --max_chars = (user input max chars/line)*2+1.
Then I use textwrap to rewrap subtitles over the max characters/line into two lines. textwrarp breaks the lines according to the following priorities:
a- Punctuation
b- before conjunctions and prepositions
c- At the last space within max characters/line
There's still improvement to be done, but so far the results are pretty good.