Every complaint about live caption accuracy is, on inspection, the same complaint: it got the names wrong. The speech itself came through fine. What came out mangled was the sponsor, the drug, the acronym, the surname of the person on stage.
That is not a model problem and it will not be fixed by choosing a different provider. It is a preparation problem, and the preparation takes about ten minutes.
Why names are the hard part
A speech recognition model is very good at words it has heard a great deal and much worse at words it has barely heard at all. “Quarterly” is everywhere in its training; your company’s internal name for a process is nowhere. Worse, the model will not return nothing. It will return the nearest common word, confidently, which is how a surname becomes a verb in front of the room.
Translation compounds it. A proper noun that survived transcription can still be translated on the way out, which is how a product called Swift becomes “rápido” in Spanish.
What a custom vocabulary does
It is a list you type in before the event, and it is used twice:
- At transcription. The terms are passed to the speech recogniser as key terms, which raises their probability against the acoustically similar common word.
- At translation. The same list acts as a do-not-translate instruction, so a brand stays a brand in all 58 languages.
Two jobs, one list. And because it belongs to the person rather than the machine, it travels with the technician between rooms and between laptops.
What to put in it
In rough order of payback:
- Every speaker’s name, spelled as it should appear. This is the one the audience notices.
- The organisation’s own name and its sub-brands. Including the one that is also an ordinary word, which is the one that will break.
- Acronyms you will hear spoken as letters, and the ones spoken as words.
- Product names. Especially the ones that mean something in the source language.
- Sector jargon. Drug names at a medical congress, part numbers at an industrial one, the regulation everybody cites by number.
- Place names the event keeps returning to.
What not to put in it: ordinary words the model already knows. A long list is not a better list; every entry you add is a probability you are nudging, and nudging “meeting” upward buys you nothing.
Get the list from the people who have it
You do not have to invent it. The list already exists, in three places:
- the running order, which has every speaker’s name;
- the sponsor deck, which has every brand exactly as legal wants it written;
- the abstracts or slides, which are where the jargon lives.
Ten minutes with those three documents is the single highest-return preparation task in live captioning, and it is the only lever that works before anything goes wrong.
Why correcting live is not the answer
It is tempting to imagine a technician fixing a wrong word as it appears. The arithmetic does not allow it: an untranslated caption reaches the screen in well under a second and a half, and a translated one inside two and a half. Nobody reads a flagged word, decides it is wrong and retypes it inside that window. The correction lands on the next line, which is a different thing and much less useful.
The operator’s live transcript, with low-confidence words flagged, is genuinely useful. It tells you how your event is going out while it goes out. But it is a monitor, not a safety net. The safety net is the list you typed before you started.