A summary nobody checked can be wrong about a decision and still read as though someone wrote it carefully. Language tooling in RoomHex is arranged around that risk. Machine output arrives marked as a draft, a named person is responsible for reading it against the transcript, and it reaches people only after that person approves it. What happens live and what happens after the call ends are separate pieces of the same job.
Stage one: while the call is running
A line of captions appears under the video as people speak. Each viewer picks their own caption language from their own controls, and that choice applies to their screen only. Two people in the same meeting can be reading two different languages while a third reads none.
The point of this stage is immediate: someone who missed a fast sentence reads it, and someone working in their second language keeps up with a discussion moving quickly.
Stage two: when the call ends
The same audio becomes a transcript with speaker labels. Each passage is attributed to whoever said it. A summary is then drafted from that transcript, covering what the meeting dealt with and what was decided. Both arrive marked as drafts, and that marking is functional: a draft is still unchecked, and the platform treats it accordingly.
Stage three: someone reviews the draft
A named person opens the draft. They read the summary against the transcript, correct names and terms the system misheard, and adjust the wording where a decision was recorded loosely. The correction workflow is where those edits are made. When they are satisfied, they approve it, and approval is what ends the draft state. Until then the draft sits with one named person and nobody else.
The reason the review stage exists is that machine output varies. Results shift with audio quality, with accents, with how technical the subject is, and with the language pair involved. What a particular meeting produced is established by reading the draft against the transcript, not by an accuracy figure.
Stage four: publication
Once approved, the transcript and summary go to the group that was already entitled to the conversation. A participant who was there opens the summary to confirm what was agreed, and someone who was invited and missed it reads it to catch up.
Where the work happens, and whose material it touches
Publication is also the moment the record leaves one person's hands. Until then a named reviewer holds it and can still change it; afterwards it belongs to the group and a change means a correction anyone can see.
Where the processing happens
The audio and text of a meeting are handled by a self-hosted service. The engine behind the work can be swapped for another one without changing where the work is done.
When the material belongs to someone else
When language tools are pointed at media that belongs to someone else, the owner's rights apply to the output as well. Translation and any dubbing are produced only where the owner's rights and policy allow it. This matters most where language tooling meets shared viewing, so that a translated subtitle or a dubbed track does not become a way around the terms under which a title may be shown.
What to decide first
- ai
- captions
- translation