Research
Can a model read brush-drawn Korean sound effects?
- Measured
- October 2026 (logged 2026-10-09)
- Sample
- 3 sound effects, 3 runs, 9 models
- Data
- 1 CSV file, CC BY 4.0
Only with the scene. In our October 2026 test of 3 brush-drawn Korean sound effects, read 3 times each, the best model read 6 of 9 from the cropped letters alone. With the scene around the word and a glossary hint, two models read 9 of 9. A larger model read 1 of 9 from the crop alone.
Table of contents
What did we test?
A comic sound effect is drawn, not typed. The strokes lean, stretch and cross the art, and one word can stand for many sounds. Before Bublate can redraw a sound effect in another language, it has to read it. We wanted to know which models read brush-drawn Korean sound effects, and what input they need.
How did we test it?
- Sample. 3 brush-drawn Korean sound effects from a copyrighted Korean webtoon chapter that we use for tests. We publish no images or text from it.
- Runs. Each model read each sound effect 3 times, so a model could score 0 to 9.
- Inputs. Three kinds: the cropped sound effect only, the crop with a glossary of common Korean sound effects, and the crop with the glossary and the scene around it.
- Scoring. We marked each read right or wrong against the word drawn on the page.
- Models. 9 models on the crop-only input. Three of them also ran the other two inputs. A tenth model was rate-limited by its provider and has no score.
Which models read them?
| Model | Crop only | Crop and glossary | Crop, scene and glossary |
|---|---|---|---|
| GPT-6 Luna (OpenAI) | 6 / 9 | 5 / 9 | 9 / 9 ($0.0005 a run) |
| Gemini 3.8 Flash (Google) | 3 / 9 | 6 / 9 | 9 / 9 ($0.008 a run) |
| Claude Haiku 5.5 (Anthropic) | 2 / 9 | 6 / 9 | 5 / 9 |
| Gemini 3.5 Flash-Lite (Google) | 2 / 9 | not run | not run |
| Claude Sonnet 5.5 (Anthropic) | 1 / 9 | not run | not run |
| DeepSeek V4.1 Flash, Qwen 3.8 Max, GLM 5.3 Flash, GPT-5.6 Luna | mostly no answer | not run | not run |
| Qwen 3.8 Flash | rate-limited, no score | not run | not run |
What do the results show?
- The scene settles the word. The same word can be a door, a punch or an explosion. With the scene, two models went from 6 and 3 correct reads to 9 of 9.
- A larger model did not read better. Claude Sonnet 5.5 read 1 of 9 from the crop alone. GPT-6 Luna, a small, fast model, read 6 of 9.
- A glossary alone is not enough. The glossary lifted Gemini 3.8 Flash from 3 to 6 and Claude Haiku 5.5 from 2 to 6. GPT-6 Luna fell from 6 to 5 with the glossary and no scene.
- The cheapest winner. GPT-6 Luna cost about $0.0005 a run with the scene, against $0.008 for Gemini 3.8 Flash.
Bublate now reads every sound effect from a crop, its scene and a glossary hint. The full story, with tips for reading a sound effect yourself, is in our blog post on this test. The meanings of common words are in the Korean sound effect dictionary.
Download the data
Each file is a CSV with a header row, comma separated, UTF-8. Measured: October 2026 (logged 2026-10-09). The files hold numbers and descriptions only, no art or text from the test chapters.
Reads per model and input
One row per model and input: correct reads out of 9, and the cost per run where we logged it.
sfx-reading-2026-10.csv · 20 rows · CSV
Columns
| Column | Meaning |
|---|---|
model | Model id on OpenRouter, or the model name where the log kept only the name |
input | crop_only, crop_glossary or crop_scene_glossary |
correct_reads | Reads marked right, empty when not scored |
reads | Reads made (3 sound effects, 3 runs each), empty when not scored |
cost_per_run_usd | Model cost of one run in US dollars, where logged |
note | Why a row has no score |
Licence and how to cite
The data is free to reuse under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence. You can copy, change and publish it, also for commercial use, if you credit Bublate and link to this page. Suggested citation:
Bublate team (2026). Brush-drawn Korean sound effect reading test (October 2026). Bublate. https://bublate.com/research/brush-sound-effect-reading. Data licensed CC BY 4.0.First published . If we correct or extend the data, the files and this page change together and the updated date above moves.
Frequently asked questions
Which model read brush-drawn Korean sound effects best?
From the crop alone, GPT-6 Luna read 6 of 9. With the scene and a glossary hint, GPT-6 Luna and Gemini 3.8 Flash both read 9 of 9. GPT-6 Luna cost about $0.0005 a run, against $0.008 for Gemini 3.8 Flash.
Why does the scene matter so much?
A sound effect word such as 쾅 can stand for a slammed door, a punch or an explosion. Brush strokes also lean and cross the art. The scene tells the model what happens in the panel, and that narrows the word down before it reads a single stroke.
Can I reuse this data?
Yes. The CSV is under the Creative Commons Attribution 4.0 licence. You can copy, change and publish it, also for commercial use, if you credit Bublate and link to this page. The suggested citation is in the section on licence and citation.
Why are there no page images?
The test chapter is a copyrighted webtoon that we use only for tests. We publish the numbers and a description of the method, not the art or the text of the chapter.