WCAG 1.4.4 resize text: do burned-in captions need it?

The answer

Burned-in captions are exempt from the resize requirement, but that does not make the size a free choice. The W3C Understanding page for 1.4.4 (read 2026-10-04) is a Level AA criterion with an exception for captions and images of text. Captions burned into pixels cannot be scaled by the viewer's browser, so the criterion does not ask for it.

What it does not remove is the practical question: a caption nobody can read does not help. That is where the size controls in the Sume video captions docs matter.

Where the size knobs live

The caption design object groups its overrides into colors, typography, placement, phrasing and motion. The typography group lists base_weight, active_weight, active_scale, font_size_ratio, safe_width_ratio and stroke_width_px. Numbers outside their documented range return a 400, so a bad value fails at request time instead of rendering wrong and billing.

One limit to remember: design is not supported on the punch or tiktok-green styles, which render on a path that reads none of those tokens. If you need to change size, pick another style, such as slam, or accept the preset.

Choosing a size for readability

Since the viewer cannot zoom burned-in text, test what you can control.

• Check the caption on the smallest screen your audience uses, not on your editing monitor.
• Keep safe_width_ratio wide enough that lines do not clip at the frame edge; the reels caption sizing post walks through that.
• Check contrast as well; the contrast checker post shows a script.
• Offer a transcript or a player with soft captions where you control the page, so readers who need larger text have it.

Open versus closed in practice

The exemption is one more reason to treat captions as a design decision, not just a compliance box.

Putting it together

Render captions with a style that supports design if size matters, set typography values within the documented ranges, and keep the script text on the page as ordinary text, which the viewer can zoom. That splits the job cleanly: pixels for people watching, text for people reading. A restyle later needs no second transcription when you pass source_caption_id, so trying a larger size costs a render, not a new speech-to-text pass.