Why Korean Subtitle OCR Loses Spaces — and How to Test It

TL;DR: Korean subtitle OCR can recognize every visible Hangul character and still return the wrong text because spaces have no visible strokes. A useful Korean OCR test must measure word boundaries separately from character accuracy. GeekLink already extracts burned-in Korean subtitles locally, and we are preparing a Korean OCR fine-tune focused on spacing for the coming week.

Why do spaces matter in Korean subtitles?

Consider this subtitle: 오늘은 정말 좋은 날이에요. An OCR result might be 오늘은정말좋은날이에요. Every Hangul syllable is present, but the line is slower to read and less useful as an editable subtitle.

The National Institute of Korean Language says in Article 2 of the Korean orthography rules that words in a sentence should, in principle, be separated by spaces. Spacing is therefore part of correctly written Korean, not optional decoration. See the official Korean language rules.

Missing spaces also affect search and replace, translation quality, names, particles, and the amount of manual cleanup a subtitle editor needs to do.

Why does Korean OCR lose spaces?

A space contains no ink, so the OCR system has to infer a word boundary from the horizontal gap between syllable blocks. That inference becomes difficult in real video subtitles.

• The model recognizes characters better than boundaries. It may have learned Hangul shapes without seeing enough accurately spaced Korean subtitle lines.
• Subtitle styling changes the gap. Outlines, shadows, condensed fonts, tracking, and compression can make a word gap look like ordinary character spacing.
• Detection can split one line into several boxes. If the joining step ignores the positions of those boxes, correct words become one concatenated string.
• Cleanup can normalize whitespace away. A valid gap may disappear after recognition during line merging or text cleanup.

Why a generic OCR accuracy score is not enough

A character score can make 다음주에 다시만나요 look almost perfect even though two word boundaries are missing. Korean subtitle OCR should report at least three separate measurements:

• Space-boundary F1: whether required spaces were preserved without inserting extra ones.
• Exact-spacing line accuracy: the percentage of lines whose spacing exactly matches the reference.
• Character accuracy with spaces removed: whether better spacing preserved the original Hangul recognition quality.

For a fair benchmark, training and test material should come from different videos. Randomly splitting frames from one episode can overstate quality because the font, outline, size, and background remain almost identical.

What is GeekLink preparing?

GeekLink is preparing a Korean subtitle OCR fine-tune focused specifically on word spacing, with the first tuning pass planned for the coming week. The goal is not to claim that every subtitle will be perfect. The goal is to make spacing a visible, measurable quality target instead of hiding it inside one headline accuracy number.

The model update will be checked with spacing metrics and with character accuracy measured separately. Until that evaluation is complete, this article describes the problem and the test design—not a finished accuracy claim.

A practical workflow for extracting Korean subtitles

If Korean subtitles are burned into the video image, they are not a normal subtitle track. OCR is needed to turn them back into editable text:

1. Import the video and select Korean as the recognition language.
2. Mark the area where the subtitles appear so logos and watermarks are ignored.
3. Run OCR across the video.
4. Review word spacing beside the original video frame.
5. Export the corrected result as SRT or TXT.

GeekLink runs this workflow locally on your computer. You can extract Korean burned-in subtitles, review each line, correct spacing, and export an editable, timestamped subtitle file.

FAQ

Disclosure: GeekLink is our own product. The spacing-focused Korean OCR model described here is in preparation; the article does not claim completed benchmark results.