Glossary
Speech bubble detection
Speech bubble detection finds the speech bubbles and text boxes on a comic page, usually with an image model that draws an outline around each one. A manga translator needs it to know where the dialogue is and how much space the translation has.
What gets detected
Speech bubble detection marks the places on a page where dialogue is written. A detection model looks at the whole page and marks the speech bubbles, thought bubbles, narration boxes and captions. Some models also find loose text written straight onto the art, which is common in webtoons. Each detected region is passed on to text recognition.
Why the bubble matters
The bubble does two jobs. It tells the program where the words are, and it defines the space the translation must fit into later. Without the bubble outline, the program would not know how far it can enlarge or wrap the new text.
Detection and reading order
Once the bubbles are found, they also have to be put in order, so that lines are translated with the right context. Japanese manga is read from right to left and top to bottom, while manhwa and most manhua are read from left to right. Because each bubble is translated in place, the page keeps its own reading order.
Where detection fails
- Faint, tiny or heavily compressed lettering.
- Bubbles with unusual shapes or no outline.
- Text outside any bubble, which many tools handle less well.
- Pages that were cropped or stitched together.
How RuneManga handles it
RuneManga finds speech bubbles on each page before reading and translating them. Text outside speech bubbles, such as narration boxes and loose text on the art, is not translated yet, and neither are sound effects. A clean, large scan with whole bubbles gives detection the best chance.
Common problems and fixes
- A whole bubble is skipped: check that it is not cut off at the edge of the image, and try a larger scan.
- Bubbles that touch or overlap are harder to separate: check that each line ended up in the right bubble.
- Text on the art is left alone: that is the known gap for text outside bubbles, not a detection fault.
Related terms
Detection is the first step. OCR reads what it finds, machine translation changes the language, and typesetting uses the bubble outline to fit the result.
Why are narration boxes not translated?
RuneManga handles text in speech bubbles today. Narration boxes and loose text on the art are the next piece being worked on.
Does the reading direction change detection?
No. Bubbles are found the same way on every page, and each is translated where it is.