Concepts in plain words¶

Who this is for: everyone — the shared vocabulary used across the wiki.

The rest of this wiki uses a handful of words again and again. Here they are without jargon. If you only read one page, read this one.

The two services¶

Speech service (sometimes "the engine" or "Whisper"): The program that listens to audio and writes it down. It is the only part that needs a reasonably capable computer. Think of it as a very fast typist who listens to the microphone.
Relay: A small message forwarder. It does not listen to speech; it just passes caption text from one browser to another. Think of it as a walkie-talkie base station: people tune in to a channel and hear what is sent.

You can use the speech service on its own. The relay is optional and only matters when you want to share captions with someone else's browser.

Capture page: The page where you pick a microphone, press Start, and see captions. Usually served by the speech service itself.
Editor: A page where a human can watch the live captions and correct mistakes before they are shown. Optional, but valuable when names and exact wording matter.
Overlay: A page that shows only the caption text and nothing else, designed to be dropped into OBS, a stream, or a projector. See Overlays for your audience.

Room: A named channel on the relay. Captions in the source room are what the speaker produces; captions in an output room are the finished, possibly corrected text. Different rooms do not mix.
Room token: A password for one room, with either read or write permission. A read token can watch a room; a write token can publish into it. The relay rejects anyone without the right token.
Service access token (CAPTION_API_KEY): A single password that protects the speech service's own API (how one program talks to another). If it is set, other programs must include it. It is separate from room tokens.

Keep tokens private

Tokens are the keys to your captions. Do not paste them into chat, screenshots or public pages. The admin pages explain how to store and rotate them.

Model: The trained data file the speech service uses. Larger models are more accurate but slower and need more memory. See Hardware and model quality.
GPU: A graphics card, used to speed up speech recognition. Strongly recommended for live use, but not required.
Origin: The web address a page is loaded from, such as https://caption.example.org. The relay only accepts connections from origins you explicitly list, which is what keeps strangers out.

Still unsure about a word? See the glossary.