What you are looking at: Caption Glass's own caption bar, playing a conversation somebody wrote by hand. This page is not listening to anything, and it does no speech recognition at all. It shows you how captions look, wrap, label who is talking and roll on. How accurate they are and how fast they arrive depend on the speech model and your PC, they are measured separately, and no page like this one can show them to you.
Below is Caption Glass's own caption renderer, the same code the app draws with, replaying a written conversation in your browser. Your microphone is not used, nothing is recorded, and no speech recognition happens on this page: the words are written into the page's source. On a Windows PC the same bar sits over your game or call and the words come from a speech model running on your own computer.
A scene drawn in code, standing in for a game. Captions are live below.
Every style in the list is one of the app's built-in styles, and every one of them meets the WCAG 2.2 AA contrast ratio of 4.5:1 for both speakers over any backdrop. Your own microphone is labeled "You"; everything the PC plays is labeled "PC audio".
What this page is, and what it is not
It is the real renderer: the layout, wrapping, speaker labels, colors, fade and roll-up you see here are the app's own captions-core.js and captions-view.js, unmodified.
It is not the speech engine. Accuracy and lag depend on the speech model and your PC, not on this page. The measured figures, the audio they were measured on and the method are in the repository's docs/MEASUREMENTS.md.
The conversation is written; no recording of any person was used to make this page.
The scene behind the captions is drawn with CSS gradients. It is not footage of any game.