How to Extract Subtitles and Ignore Foreign-Language Logos & Watermarks

Key takeaways

• Re-uploaded and re-posted videos often carry a station logo or watermark in a different language than the dialogue — a Thai, Korean, or Japanese platform logo on a Chinese clip, for example — and ordinary OCR reads all of it into your SRT.
• GeekLink filters OCR by writing system: set your subtitle language and it keeps text in that language's script (plus Latin letters and numbers) and drops text in other scripts automatically.
• This catches noise that color and size filters can't — a foreign-language logo is removed even when it shares the subtitle's color and size, because it's simply a different script.
• It does not remove English. Latin letters and digits are always kept, because real subtitles routinely mix in English words and numbers.
• It runs offline on Apple Silicon Macs across 90+ languages and exports a clean SRT.

Why does foreign-language text end up in my extracted subtitles?

OCR reads every piece of text in the frame, regardless of language. When a clip has been re-uploaded or re-posted, it frequently carries the original platform's watermark or a station logo — and that text is often in a completely different language from the spoken dialogue. OCR can't tell which language is "the subtitle"; it just reads characters.

This is extremely common for short-form and re-posted video. A Chinese drama clip re-uploaded from another platform might show a Thai or Korean station bug in the corner; a compilation might carry a Japanese channel name across the top. Read literally, all of that lands in your SRT next to the real dialogue.

The reliable way to separate them is by writing system. If your dialogue is Chinese and the logo is Thai, the two use entirely different scripts — so telling the OCR engine "my subtitles are Chinese" gives it a clear rule for what to keep and what to discard.

How do I filter OCR by language?

You set the subtitle language before extraction, and GeekLink keeps only text whose writing system matches that language — plus Latin letters and numbers — and drops everything else. The language you choose does two jobs: it selects the recognition model tuned for that language, and it defines the set of "expected" scripts for the output.

When a detected line is in a script that doesn't belong to your language — Japanese kana on a Chinese video, Thai on a non-Thai video — GeekLink treats it as not-your-subtitle and leaves it out. You don't toggle anything extra; choosing the subtitle language is the filter.

Because this works on the script, not the appearance, it removes foreign-language noise that color and size filters miss — a logo that happens to be white and subtitle-sized still gets dropped if it's written in another language.

What does each language keep and drop?

Each subtitle language keeps its own script plus Latin (for the English and numbers that appear in almost any subtitle), and drops unrelated scripts. Here is how the common languages behave:

Subtitle language	Kept	Dropped (examples)
Chinese	Chinese characters + Latin/numbers	Japanese kana, Korean, Thai, Cyrillic, Arabic
Japanese	Kana + Chinese characters + Latin/numbers	Korean, Thai, Cyrillic, Arabic
Korean	Hangul + Latin/numbers	Japanese kana, Thai, Cyrillic, Arabic
Thai	Thai + Latin/numbers	CJK, Korean, Cyrillic, Arabic
Russian	Cyrillic + Latin/numbers	CJK, Korean, Thai, Arabic
Arabic	Arabic + Latin/numbers	CJK, Korean, Thai, Cyrillic
Hindi	Devanagari + Latin/numbers	CJK, Korean, Thai, Cyrillic

For a Latin-script subtitle language (English, Spanish, Portuguese, and so on), the expected set is Latin — so non-Latin logos and watermarks are dropped while your dialogue is kept.

Does language filtering remove English text too?

No. Latin letters and numbers are always kept, no matter which subtitle language you choose. Subtitles routinely mix in English words, brand names, and digits — a Chinese line might include "OK," a model number, or a year — so stripping Latin would damage real subtitles.

That means language filtering removes other non-Latin scripts (Japanese, Korean, Thai, Cyrillic, Arabic, and so on), not English. If your noise is an English watermark on an English-subtitle video, use the color, font-size, and region filters instead — those separate same-language text by appearance and position.

What about bilingual subtitles, like Chinese with English?

Bilingual subtitles are supported — GeekLink can capture both languages of a dual-language subtitle. Many Chinese videos carry Chinese on one line and English on the line below; because Latin is always kept alongside Chinese, both lines come through.

This is different from dropping a foreign-language logo: a bilingual subtitle is intentional dialogue in two scripts that belong together, while a logo is unrelated text in a third script. Language filtering keeps the former and drops the latter.

Step by step: drop the foreign logo and keep your subtitles

1. Import your video(s). GeekLink batch-processes a whole folder at once.
2. Set the subtitle language to the language of your dialogue (for example, Chinese). This defines which scripts are kept.
3. Choose the subtitle region and start OCR. Pick where the subtitles sit; then pick the subtitle color from sample frames if you also want to filter by appearance.
4. Let it run. GeekLink reads the dialogue and automatically discards lines in other writing systems — the foreign station logo or watermark never reaches your SRT.
5. Export SRT. You get a clean file with only your subtitles, ready to translate or edit.

Stack it with the appearance filters for the cleanest result: language drops the foreign-script logo, while color, size, and region clean up any same-language noise. See the full guide on extracting subtitles without the watermark, or the OCR guide in the docs for a reference on every setting.

Why this matters for re-uploaded and re-posted video

If you work with re-posted clips, foreign-language logos are the most common source of junk in extracted subtitles. Creators who pull dialogue from short dramas, compilations, and cross-platform re-uploads constantly deal with station bugs and platform watermarks in Thai, Korean, or Japanese sitting on top of Chinese dialogue.

Filtering by language turns that from a manual clean-up job into an automatic one: set the dialogue language, and the foreign text is gone before you ever open the SRT. Combined with batch processing, you can clean a folder of re-posted clips in a single pass — all offline on your Mac.

Disclosure: GeekLink is our product. The language/script behavior described here reflects how the app filters OCR output. Always confirm results on your own footage, since unusual fonts and stylized logos can vary.

Example: a Chinese clip with a Thai station logo

Say you're re-posting a Chinese short-drama clip that still carries a Thai platform's logo in the top-right corner and a small English timestamp at the bottom. Here is what each filter does:

• Set the subtitle language to Chinese. The Thai logo is in a different script, so it is dropped automatically — even though it sits over the video like the subtitles do, and even if it is white and similarly sized.
• The English timestamp stays by default, because Latin letters and numbers are always kept. If you don't want it, add a region filter (it sits at the bottom edge, outside the subtitle band) or a color filter (it is usually a different shade).
• The Chinese dialogue comes through clean, ready to export as SRT or translate to English for a bilingual version.

The outcome is identical whether you process one clip or fifty — set the language once and batch the whole folder, offline, in a single pass.

How is language filtering different from filtering by color or size?

Language filtering separates text by writing system; color, size, and region filters separate text by how it looks and where it sits. They solve different problems, and the cleanest extractions use both together.

Reach for the language filter when the noise is in a different language from your dialogue — a foreign station logo or platform watermark. It removes that text no matter what color or size it is, because it's simply a different script. This is the only filter that can drop a logo which happens to match your subtitle's white color and exact height.

Reach for the color, size, and region filters when the noise is in the same language as your dialogue — a Chinese channel name on a Chinese video, for instance. Language filtering can't tell those apart, but appearance can: the watermark is usually a different color, smaller, or parked in a corner.

In practice, set the language first to clear out foreign scripts, then add color and size to clean up any same-language noise that remains. The two passes are complementary, not redundant.

What if some foreign-language text still gets through?

A few logos are designed to be hard to read, and those edge cases are worth knowing. Heavily stylized or decorative logos can be misread by OCR as random Latin characters rather than their real script — and because Latin is always kept, that garbled output can occasionally slip into the SRT.

When that happens, layer on the appearance filters: a corner logo is removed by restricting the OCR region, and a faint or off-color bug is removed by the color filter. Between language, region, and color, almost every logo falls to at least one of them.

Lines that genuinely mix two foreign scripts are rare in real subtitles, but if you hit one, pick the language that matches your dialogue and let the appearance filters handle the rest. The goal isn't one perfect rule — it's stacking a few cheap filters so the leftover clean-up is trivial instead of line-by-line.

FAQ