New args: sentence, standard, max_line_width, max_line_count, max_comma, max_gap.

[Purfview]

Post feedback about these here.

New parameters :
--sentence
--standard
--max_comma
--max_gap
--max_line_width
--max_line_count

Implemented since r160.8.

r160.9 UPDATE:

Bugfix: The final word could disappear in some combination of the new parameters from r160.8.
New parameter : --standard_asia

r160.10 UPDATE:

Bugfix: The final line disappeared if the final word wasn't punctuated when using --sentence.

[BenJamesAndo]

Thanks for adding these! I'm now able to get very close to where I'd like to get.
Firstly a bit of feedback on the help file. It says under --sentence
... Be default meant to output whole sentence per line for better translations, but not limited to. (default: False)

I'm not quite sure what that means. Is it supposed to say something like: The default behaviour is to add a whole sentence per line for better translations, but not be strictly limited by it?

If I run a command with --max_line_width 37 --max_line_count 2 then I get

Error: Original functionality of 'max_line_width' and 'max_line_count' parameters is borked.
Only the alternative funtionality is allowed with --sentence.

Is this saying that --sentence is mandatory when using --max_line_width & --max_line_count? Because it doesn't say that it's required in the help file.

I've tested this out on a few 28 minute videos. While it technically works flawlessly, it does lead to another problem. Sometimes I have a line with only one word in it.

I wonder if it's worth having a new parameter that specifies the minimum amount of characters on a line. e.g. --min_line_width 8. It could get quite complicated as it may not always be able to fulfil the requirements. But in the above examples, it could certainly move a few words from the first section into the section with only one word.

[BenJamesAndo]

So do you think for my situation I'm best to just stick with --one_word and then use Subtitle Edit to merge the lines together? Or is there some other way to avoid having only one word in a line, whilst still adhering to a certain max line width and max line count?

[Purfview]

You say it like SE has min_line_width option or like you wouldn't get those one words there too.

min_line_width would break many other things. Tell me this magic what to do if min_line_width would be met in your examples.

[BenJamesAndo]

Sorry I misread your response SubtitleEdit/subtitleedit#7655 (comment)
I thought you were implying that these new options would help with the one word on a line problem.
You are right, using SE via CLI also produces the same result.

Intererestingly though, generating captions with Faster-Whisper without any arguments directly in the SE GUI, it actually does somehow adhere to the single line max. length. that's set in the general settings.
It's because of having Auto adjust timings option enabled, which doesn't seem accessible via the CLI. I haven't done extensive testing, but it does seem to do quite a good job. I'm just trying to get the same results through only the CLI.

[Purfview]

I thought you were implying that these new options would help with the one word on a line problem.

That thread is about merging --one_word, which is the option from Standalone Faster-Whisper, not about "one word problem".

Intererestingly though, generating captions with Faster-Whisper without any arguments directly in the SE GUI...

I think that's because of "post-process" option is selected. I personally deselect all there as it does something weird. I think those were meant for some poor whisper implementations, I recommend to disable them all with Standalone Faster-Whisper.

[oep42]

@BenJamesAndo I suggested a tool to add to Subtitle Edit, to automatically solve the "short line" or "one word on a line" problem after subtitles have been created with a Whisper engine. I hope that a tool like this will be added to Subtitle Edit. See here.

[ext666]

Looks great, thanks.

Would it be possible to make a command that would limit the maximum length of the subtitle in seconds? Something like max_time 6 (for 6 seconds).

[Purfview]

Would it be possible to make a command that would limit the maximum length of the subtitle in seconds? Something like max_time 6 (for 6 seconds).

Possible, what settings do you use and could you provide json file for example?

[despairTK]

Hi! After I tried the new parameter --sentence, I found it very useful. But I also have some doubts.

For example:

337
00:34:27,020 --> 00:34:40,420
What I want you to focus on and what I'm trying to focus on with this particular study is a breakdown of the lights and the darks and an indication of these things with very, very simple shapes because I'm going to start moving forward on the painting now,

338
00:34:40,480 --> 00:34:43,980
and I want you to look at how I proceed.

Before adjustment:
[34:27.020 --> 34:31.640] What I want you to focus on and what I'm trying to focus on with this particular study is
Processing segment at 31:01.260
[34:32.040 --> 34:37.180] a breakdown of the lights and the darks and an indication of these things with very, very
[34:37.180 --> 34:40.940] simple shapes because I'm going to start moving forward on the painting now, and I want you
[34:40.940 --> 34:43.980] to look at how I proceed.

This is also true in another place:

340
00:34:51,120 --> 00:35:12,220
I'm going to kind of average the skin tone of his face and start there, and I want you to see how quickly, because I have these shadow shapes established and I have the drawing in place done by putting these value shapes in their place , I can actually focus wholly now on the transition areas,

341
00:35:12,320 --> 00:35:16,080
and also I can focus on putting really expressive brushwork in here.

Before adjustment:
[34:51.120 --> 34:57.340] I'm going to kind of average the skin tone of his face and start there, and I want you
Processing segment at 31:26.960
[34:57.340 --> 35:02.720] to see how quickly, because I have these shadow shapes established and I have the
[35:02.720 --> 35:08.920] drawing in place done by putting these value shapes in their place, I can actually focus
[35:08.920 --> 35:15.680] wholly now on the transition areas, and also I can focus on putting really expressive brushwork
[35:15.680 --> 35:16.080] in here.

The above two examples do not form a complete sentence, but are separated by commas into two lines of subtitles. Although it does not affect translation, sentences with more words than them are well segmented. Don't know if this is normal.

[Purfview]

"Wonderful Stylized Backgrounds and Planar" is 41 chars, + "101" would be over 42 chars.
Isn't "101" in the next line?
Your screenshot doesn't show what is after 32 line, stop confusing me. 😆

[despairTK]

it's here
01.zip

To be honest, I will read the --help information. However, since machine translation cannot accurately translate some --help information, I can only use audio testing to understand what each parameter does. . . .

Add a screenshot and audio, you can also test it.
Audio.zip

[Purfview]

Yeah it's the bug, can happen only with last word in the whole file.

[Purfview]

Test if it's fixed in r160.9

[despairTK]

After testing, this problem was indeed fixed.

[Purfview]

I think new parameter for comma should be introduced, for example if comma is near at the line's end then it should break the new line, I meant don't add one or two short words after it.

[1001ruchka]

Tested the operation with the - "--sentence" parameter.
When recognizing speech for the same video. When using the parameter "--sentence", the last sentence in the video was not recognized.
Without the parameter "--sentence" it was recognized without any problems..

[Purfview]

I didn't understand the question about which txt file you are talking about, I use Faster-Whisper in SE.

One of the outputs formats. With -f=all you can output them all. Share srt and json files.
I'm not interested in what you see in SE nor in audio file.

[1001ruchka]

Purfview-Whisper-Faster.zip
archive with all files

[Purfview]

Thx, now I see that it's some bug. I'll look at the code.

[Purfview]

Should be fixed in r160.10, hopefully no more bugs. 🪲

Good way to check:
With SE open .srt and export as "Plain text" and take .text file, now we can test contents of both files in https://wordcounter.net
Word count should match, otherwise - Houston we have a problem. 😆

[Purfview]

Btw, I accidentally noticed wrong line split in your file at 19 sub, the split at Dr..
I made ignore list from SE lists, it will be in the next release:

ignore_list = ['Dr.', 'Mr.', 'Mrs.', 'Ms.', 'Frk.', 'Hr.', 'Sra.', 'Srta.', 'Dra.', 'Prof.', 'Mtro.', 'Mtra.', 'Lic.', 'Arq.', 'Ing.', 'Cnel.', 'Tte.', 'Gral.', 'dipl.', 'bl.', 'gosp.', 'ing.', 'mr.', 'prof.', 'sv.', 'vlč.', 'dr.', 'g.', 'DR.', 'DRA.', 'eng.', 'Eng.', 'exmo.', 'Exmo.', 'gov.', 'Gov.', 'jr.', 'Jr.', 'mna.', 'Mna.', 'mno.', 'Mno.', 'mrs.', 'P.', 'pe.', 'Pe.', 'srs.', 'Srs.', 'St.', 'vol.', 'vols.', 'sr.', 'Sr.']

[1001ruchka]

Please clarify the meaning of the parameters:
--standard
--max_comma
--max_gap
unfortunately I could not find a description

[Purfview]

--help

[despairTK]

In version r160.10, a sentence segmentation error occurs when the --sentence parameter is enabled. as the picture shows:

Here are all the files output.
test.zip

My transcription settings：--model large-v2 --beam_size 10 --best_of 10 --sentence --compute_type=float16 --max_comma 1000

[Purfview]

Not error, there is over 3.0 seconds gap between There and you go..
max_gap default is 3.0.

Why there are such big gaps between words? Are you doing a translation here?

I feel has 10 seconds gap.
All right has 80 seconds gap. 😄

[despairTK]

The problem is that the original normal sentence is strangely split after using the --sentence parameter.

[Purfview]

The problem is that the original normal sentence is strangely split after using the --sentence parameter.

It's split correctly according to the settings.

It's not "strangely split", timestamps are "strange" in your transcription:

[despairTK]

terribly sorry. . . I didn't check carefully. It's a waste of your precious time. . . You can delete this feedback conversation as no real feedback was provided.

[Purfview]

Nah, let it be. Maybe I'll clean up the images.

Maybe it's not clear but:
--sentence activates all those new max_.... parameters.

[Azzhoe]

Post feedback about these here.

New parameters : --sentence --standard --max_comma --max_gap --max_line_width --max_line_count

Implemented since r160.8.

r160.9 UPDATE:

Bugfix: The final word could disappear in some combination of the new parameters from r160.8. New parameter : --standard_asia

r160.10 UPDATE:

Bugfix: The final line disappeared if the final word wasn't punctuated when using --sentence.

what does the --standard_asia do?

[oep42]

Suggestion 1: Make the --standard parameter versatile, by enabling it to pass numeric values to its "built-in" parameters, like this:

• Without a numeric value, the current hard-coded values of its "built-in" parameters are being used, so these four parameters are activated with these values: --sentence, --max_line_width 42, --max_line_count 2, --max_comma_cent 70.

• With a numeric value such as this --standard 38,3,50, these four parameters are activated with these values: --sentence, --max_line_width 38, --max_line_count 3, --max_comma_cent 50.

Suggestion 2: Make explicit in the --help information what the --standard parameter exactly does, by explicitly mentioning all four parameters that are activated by --standard. In the current description, it is not obvious to those unfamiliar with the subject, that the parameters --max_line_width and --max_line_count are meant.