Improving the subtitles in the original language with Gemini Flash 2.0?

[Timo425]

I was trying to improve the subtitles in original language, to add punctuation and fix grammar in general, but the tool just translates to English every time. I selected instructions (improve quality).txt with default settings.

Then I created a copy of that file in the original language of the series (Estonian) and also added a note in Description of project settings:
"I want to improve the grammar and punctuation in the original language (Estonian)."
The combination of these two additions finally did the trick.

Is this a quirk of Gemini Flash 2.0 or the tool doesn't work very well for improving subtitles in the same language unless you modify the instructions and/or description?

Sorry if its a bit foolish question, I'm using this program for the first time.

[machinewrapped]

Thanks for the report! I suspect that because the instructions are in English and include English examples it predisposes the model to respond in English, even when it requires translation. It's possible another model would understand the task but Gemini 2.0 Flash is pretty smart, so it's probably just a lack of clarity in the instructions.

I'll update the instructions to include some non-English examples and to make it explicit that this is not a translation task. You can edit the file Instructions (Improve Quality).txt yourself in the meantime, something like:

### instructions
The goal is to improve the quality of a set of subtitles, without changing the language (this is not a translation task).

You will receive a batch of lines and should respond with richer, natural dialogue in the same language, including appropriate punctuation, correcting any spelling or grammatical errors.

If you add an example in Estonian that will help the model understand that it shouldn't change the language, too.

In the next update I'll also make the request/response format more flexible, as it currently requires the model to respond with a "Translation" for each line even for quality improvement tasks. I'm having issues with Qt crashes (the UI framework) so I can't publish a new release just yet, but fingers crossed I'll get it done this weekend.

[machinewrapped]

Try version 1.1.0 :)
https://github.com/machinewrapped/gpt-subtrans/releases/tag/v1.1.0

Tested with Spanish and Korean subtitles using Gemini 2.5 Flash Experimental... can't say for sure the quality was improved, but they at least stayed Spanish and Korean!

[Timo425]

Thanks for the quick reply and fix!
Now, with version 1.1.0, the improving seems to keep the original language!
However, i see that the Translation> lines are now replaced with Improvement> lines and the tool is not reacting well with that?
The final subtitles end up like this:

670
00:44:41,885 --> 00:44:47,380
Original>
взрослым в определённом возрасте сейчас мы узнаем в каком его
всю ночь избивают плётками вы до крови бьёте да бьём очень сильно
Improvement>
взрослым, в определённом возрасте (сейчас мы узнаем, в каком), его
всю ночь избивают плётками. Вы до крови бьёте? Да, бьём очень сильно,

EDIT: replacing Improvement> with Translation> seems to help.
Anyway I guess this is the request/response format thing you are talking about. :)

EDIT2: nvm, some subtitle blocks were still translated.

[machinewrapped]

Hi, thanks for the quick test :)

Yes, part of the change was to replace "Translation" with "Improvement" in the response format (or whatever is specified in the task_type field in the instructions file). The response parser is supposed to look for that - sounds like it's not getting the task_type. Was this with a new project or an existing one?

[machinewrapped]

I've updated the release (same version number, just re-download) to show the task type in the "Edit Instructions" dialog. If it doesn't match the instructions, i.e. "Improvement" you can correct it by hand, or reload the instructions to refresh it.
https://github.com/machinewrapped/gpt-subtrans/releases/tag/v1.1.0

Let me know if you still have problems and see if you can tell when it goes wrong, e.g. after reloading a project :)

[Timo425]

I'm a bit scared to download the latest build, windows antivirus seems to think it has a Trojan:Script/Wacatac.B!ml

is it safe?

[machinewrapped]

Yeah, should be safe - Windows Defender sent it to the servers to be scanned when it was built and it came back clean. False positives with antivirus software are unfortunately common with programs built/packaged with PyInstaller (#231).

I just manually scanned the zip file with Defender and it didn't complain either... sometimes seems like it just rolls a dice!

[Timo425]

Okay, tried again with the new 1.1.0.
Noticed that task_type was set to Translation even though I had Improvement selected (and I even had clicked Defaults), so I manually wrote Improvement there.

Hard to say if any other leftovers from earlier were left, but I know there was no earlier project saved from this specific improvement.
Anyway, it still translates. :/

I asked Gemini Pro 2.5 to revise the instructions, so this is what it came up with:

[CRITICAL] Primary Directive: Maintain Original Language

Your only task is to improve the provided subtitle lines within their original language.

DO NOT TRANSLATE the subtitles into English or any other language. This is not a translation task.

The language of your Improvement> response MUST be the exact same language as the Original> input for that line number. If the input is Spanish, the output must be Spanish. If French, the output must be French, etc.

Goal: Subtitle Quality Enhancement (In Original Language)

Improve Grammar and Spelling: Correct any grammatical errors or misspellings found in the original line.

Add/Correct Punctuation: Ensure appropriate punctuation (commas, periods, question marks, etc.) is used for clarity and natural flow. Crucially, do not add ellipsis points (...) at the beginning or end of lines unless they are grammatically essential to represent a pause, interruption, or trailing thought that is clearly implied or present in the original dialogue. Avoid adding them purely for stylistic connection or transition between lines where standard punctuation suffices.


Enhance Naturalness: Rephrase awkward or unnatural-sounding sentences to make the dialogue sound more fluent and realistic in the original language, while preserving the original meaning. This should be achieved through better word choice and sentence structure, not by unnecessarily adding transitional punctuation like leading/trailing ellipses.

Maintain Meaning: Do not change the core meaning or intent of the original subtitle line.

Input & Output Requirements:

You will receive a batch of subtitle lines, each with a line number.

Your response MUST contain the exact same number of lines/line numbers as the input batch.

Each improved line must correspond directly to its original line number.

Using Context and Specific Rules:

Use any provided context (like previous batches or scene descriptions) to inform your improvements and ensure consistency (e.g., tone, character voice).

If a Name List is provided, strictly adhere to the spellings given in that list.

Correct obvious errors (typos, clear grammatical mistakes) using context if possible. Do not invent information or details not implied by the original text or context.

If the original line contains profanity, do not censor or remove it. Reproduce it accurately in the improved line.

Mandatory Response Format:

Your response MUST strictly follow this format for each line:


#[Line Number]
Original>
[The original subtitle line verbatim]
Improvement>
[Your improved subtitle line IN THE ORIGINAL LANGUAGE]



IGNORE_WHEN_COPYING_START

Use code with caution.
IGNORE_WHEN_COPYING_END

After processing all lines in the batch, add the following tags at the very end:


<summary>[A one or two-line synopsis IN ENGLISH of the dialogue covered in this specific batch.]</summary>
<scene>[A short summary IN ENGLISH of the current scene, incorporating information from previous batches if relevant/provided.]</scene>



IGNORE_WHEN_COPYING_START

Use code with caution.
IGNORE_WHEN_COPYING_END

(Note: The summary and scene descriptions should be in English for easier processing, even though the subtitle improvement itself is not.)

Examples (Illustrating Format and Improvement TYPE - NOT Language Task):

These examples show the required format and the type of grammatical/stylistic improvements expected. Remember, your task is to perform similar improvements BUT strictly in the language of the input you receive.

Example 1 (English Input -> English Output):


#146
Original>
Had if not been him
Improvement>
Had it not been for him,

#147
Original>
The girl would have been trampled to dead
Improvement>
the girl would have been trampled to death.



IGNORE_WHEN_COPYING_START

Use code with caution.
IGNORE_WHEN_COPYING_END

Example 2 (English Input -> English Output):


#100
Original>
How do you think she's heading
for so hurry up
Improvement>
Where do you think she's
heading in such a hurry?



IGNORE_WHEN_COPYING_START
Use code with caution.
IGNORE_WHEN_COPYING_END

Example 3 (Spanish Input -> Spanish Output):


#204
Original>
Pero yo ya hice mi decisión
Improvement>
Pero yo ya tomé mi decisión.



IGNORE_WHEN_COPYING_START
Use code with caution.
IGNORE_WHEN_COPYING_END

Example 4 (French Input -> French Output):


#310
Original>
Tu doit parler avec elle avant qu’il sera trop tard
Improvement>
Tu dois lui parler avant qu’il ne soit trop tard.



IGNORE_WHEN_COPYING_START
Use code with caution.
IGNORE_WHEN_COPYING_END

Example 5 (German Input -> German Output):


#401
Original>
Ich habe gesehen ihn gestern
Improvement>
Ich habe ihn gestern gesehen.

Seems to work!

Although, gemini flash 2.0 is not that great in improvement anyway, I've had better luck just dumping raw from srt, 300-400 subtitle blocks at at time, to gemini 2.5 pro in ai studio and it's doing it great like this (maybe I could do the same with this tool though, just need to figure out how big batches can I allow..).

On a sidenote, not related to improvement: gemini 2.0 flash also messes up summary parts sometimes in regular translating, so I made the summary thing also more explicit in the instructions (just an example to get the idea):

At the end you MUST add <summary> and <scene> tags with information about the translation:

<summary>**This should be a brief, high-level description of the main topic, action, or purpose of the dialogue in the current batch, written in your own words.** It must synthesize the content, not merely repeat it. **Crucially, do NOT include direct quotes or translations of individual lines from the batch within the summary tag.** For example, avoid putting specific dialogue fragments like "Me peaksime vähemalt küsima." or "Liv, ära isegi mõtle sellele." inside the <summary>. Instead, describe *what is happening* (e.g., "Characters debate whether to ask a question," "A character warns another against a specific action," or like the good example: "The team discusses the details of the kidnapping case, including the victims' descriptions and medical conditions."). Aim for one or two descriptive sentences.</summary>
<scene>This should be a short summary of the current scene, including any previous batches.</scene>

[machinewrapped]

Hmm, curious! Make sure the instructions file has these lines in (should be just after the ### prompt)

### task_type
Improvement

If you've got the >Improvement change you should have that part too though!

I'll take a look at Gemini's suggestions and see if I can improve the default instructions. It's good practice to create custom instructions tailored for your specific requirements though - consider translating the instructions into Estonian to make sure the model is in the right frame of mind from the start :)

[Timo425]

### task_type
Improvement

Oops, yeah I think I had overwritten that somehow. That was it. Thanks!

[Timo425]

@machinewrapped Okay, I found out how the task_type gets disappeared. Save Instructions action under Edit Instructions somehow disappears the task_type entry from the file completely (even when the task type is defined under edit instructions view).

[machinewrapped]

Ah! Good catch - yeah, I forgot to test that 🤦 Easily fixed, I'll put up a new version soon.