Can software read brush-drawn Korean sound effects?

Reading a brush-drawn 쾅 is harder than reading a speech bubble. Giving the model the whole scene mattered more than which model we used.
Yes, if it can see the scene. In our test of 3 brush-drawn Korean sound effects, read 3 times each, the best model read 6 of 9 from the cropped letters alone and 9 of 9 with the surrounding scene and a glossary hint. A large model read only 1 of 9 from the crop. Context mattered more than model size.
A comic sound effect is drawn, not typed. Translation research treats sound effects outside balloons as visual parts of the page, with their own size, shape and colour (Jagyeong Kim, inTRAlinea, 2023). Strokes overlap the art, letters lean and stretch, and one word can stand for many sounds. The Standard Korean Language Dictionary defines 쾅 as the sound of a heavy, hard object that falls or strikes something, and Wiktionary glosses it as "bang" or "boom". On a page, that one word can be a slammed door, a punch or an explosion. Before Bublate can redraw a sound effect, it has to read it.
How did we test it?
We took three brush-drawn Korean sound effects from a test chapter and asked each model to read them three times, so 9 reads per model. We tried three kinds of input: the cropped sound effect only, the crop with a glossary of common Korean sound effects, and the crop with the glossary and the surrounding scene.
Which models read them best?
| Model | Crop only | Crop + glossary | Crop + scene + glossary |
|---|---|---|---|
| GPT-6 Luna (small, fast) | 6 / 9 | 5 / 9 | 9 / 9 |
| Gemini 3.8 Flash (mid-size) | 3 / 9 | 6 / 9 | 9 / 9 |
| Claude Haiku 5.5 (small) | 2 / 9 | 6 / 9 | 5 / 9 |
| Gemini 3.5 Flash-Lite (light) | 2 / 9 | not run | not run |
| Claude Sonnet 5.5 (large) | 1 / 9 | not run | not run |
Four more models (DeepSeek V4.1 Flash, Qwen 3.8 Max, GLM 5.3 Flash and GPT-5.6 Luna) mostly returned no answer at all on the crop-only test. A tenth, Qwen 3.8 Flash, was rate-limited by its provider during the test, so it has no score. The method, the limits and the raw data as a CSV file are on the research page for this test.
What did we learn?
- Context beats model size. A large model read 1 of 9 from the crop alone. The scene told the smaller models what was happening, and that settled the reading.
- A glossary helps, but not alone. A list of common sound effects lifted weak readers, and slightly confused the best one when it had no scene to check against.
- Cost follows the winner. The model that read 9 of 9 with context cost about $0.0005 a run, a sixteenth of the next best.
The same 쾅 is a door in one panel and an explosion in the next. A model that only sees the letters cannot know which.
How do you read a brush-drawn sound effect yourself?
The test points to the same method a person uses. Read the scene first, then the strokes.
- Look at what happens in the panel. A door, a fist, a fall or a fire narrows the word down before you read a single stroke.
- Find the syllable blocks. Each Hangul syllable sits in its own block, so count the blocks first. A repeated block, as in 쾅쾅, is a repeated sound.
- Check a dictionary. Look the candidate up in a dictionary of Korean sound effects, and compare its meaning with the scene.
- Pick the English word for the scene, not for the dictionary. A 쾅 on a door is a SLAM, on a gun it may be a BANG. Our guide to translating Korean sound effects has more examples.
How does Bublate use this?
Bublate now reads every sound effect from a crop, its scene and a glossary hint, and the text detector's own reading counts as a second vote on the result. The words then go to the redraw step, explained in how we redraw sound effects. For why most other tools skip this step, read why reader tools skip sound effects.
Method: 3 sound effects, 3 runs each, Korean to English, October 2026. Raw data (CSV, CC BY 4.0) and the full method: brush sound effect reading test.
Frequently asked questions
Why is a brush-drawn sound effect hard to read?
The letters are drawn as art. Strokes lean, stretch and overlap lines and screentone, and one word can stand for many sounds. A reader that sees only the letters has no way to tell a slammed door from an explosion. The scene around the word settles it.
Does a bigger model read sound effects better?
Not in our test. A large model read 1 of 9 brush-drawn Korean sound effects from the crop alone, while a small, fast model read 6 of 9. With the scene and a glossary, the small model read 9 of 9. Context mattered more than size.
Does a glossary of common sound effects help?
It helps weak readers. In our test, a glossary lifted two models from 3 and 2 correct reads to 6. It slightly confused the best model when it had no scene to check against. The best result came from the glossary and the scene together.
Which models did you test?
We tested 9 models on the crop-only input: GPT-6 Luna, Gemini 3.8 Flash, Claude Haiku 5.5, Gemini 3.5 Flash-Lite, Claude Sonnet 5.5, DeepSeek V4.1 Flash, Qwen 3.8 Max, GLM 5.3 Flash and GPT-5.6 Luna. The first 3 also ran the other two inputs. The raw data is on our research page for this test.
How does Bublate read sound effects now?
Every sound effect is read from a crop of the letters, the scene around them and a glossary hint. The text detector's own reading counts as a second vote. A person can then check and change the word in the editor before export.
Sources
Related
Written by the Bublate team. We build tools that translate and letter comics, and we publish what we measure along the way.
Stories about comics in every language, once a month
Related articles
Translate a comic chapter from Claude or ChatGPT (MCP)
By the Bublate teamBublate runs a remote MCP server at bublate.com/mcp. Connect Claude, ChatGPT or Cursor, and ask it to translate and re-letter a webtoon or manga chapter.Comic translation API: chapters from your own pipeline
By the Bublate teamThe Bublate API takes webtoon and manga pages, translates and re-letters them, and calls your server when a chapter is ready. Five requests, with curl examples.