Read this first: the term list is effective only on target languages that are routed to the Qwen engine. Target languages routed to Gemini have no equivalent feature — there is no field to fill in, and nothing you type is used. This is a limit of the engines themselves, not a setting we have hidden. If the language you translate into runs on Gemini, the list you build here has no effect on that session, and you should plan around the engine getting a name wrong rather than around a list fixing it.
Which engine serves a given target language is decided on our side, per target language — it is not something you pick in the app. So the honest way to use this feature is: build the list, run one short session in the target language you actually use, and see whether the names come out right. If they don't change at all, your target is not on the engine that reads the list.
What a term list actually does
Speech recognition fails on words it has no reason to expect. A surname, an internal product code, a company nobody outside your industry has heard of — these come back as the nearest ordinary word, and once recognition is wrong the translation is confidently wrong too. A term list is a hint given to the engine before the audio starts: these strings are likely to occur.
Three kinds of entry are worth the space:
- Product and brand names — your own, your customer's, the tool you are demoing.
- People's names — the ones that will actually be spoken aloud in this session, not your whole address book.
- Field-specific vocabulary — the handful of terms that carry the meaning of the conversation and that a general model has no reason to prefer.
We do not publish an accuracy figure for this, and we are not going to. The effect depends on the term, the target language, the speaker's accent and how the word sits in the sentence. It is a hint, not a substitution rule: the engine can still ignore it. Judge it on your own recordings — a session's transcript is saved locally, so you can compare a run with the list against a run without it. See transcripts and history for where those files land.
Three ways a term reaches the engine
1. The list you type
The Terms window in the app holds your own list — one entry per line, edited whenever you like, kept between sessions. This is the durable part: the vocabulary that belongs to your work rather than to one meeting.
2. A file you import
A term list can be imported from a file instead of typed. That is what makes it something you can prepare elsewhere, review with a colleague, and load on a second machine without retyping it.
3. The names you enter before a meeting
Before a call, the main window has two short fields for the participants — your name and the other side's name. What you type there joins the term list for that session. It is worth the ten seconds: the two names said most often in the next half hour are the two the engine is least likely to know. This pairs with Meeting mode, where names are spoken constantly.
The limit is 200 entries — and it is a budget, not a quota
The combined list is capped at 200 entries. This is not a server restriction we could lift for a bigger plan. It is a latency and bias budget. Every term you add is something the recogniser is nudged toward, and a long list has two costs: the hint travels with the session, and a model biased toward 200 unusual strings starts hearing them in audio where they were never said.
The failure mode to watch for is over-biasing. Load a list of every product in the catalogue and you will start seeing product names appear in sentences that had nothing to do with products. A list of fifteen terms that are genuinely going to be spoken beats a list of two hundred that might be.
Treat the cap as guidance, not a target. If you are near it, the question is not "how do I get more slots" but "which of these will be said out loud today".
The shipped default list arrives switched off
Voxis ships with a small default list of common technology terms. It arrives switched off, and the app asks you once whether you want it. That is a deliberate reversal: it used to ship on, and that was a defect, not a feature. A default list is a set of assumptions about your vocabulary made by people who have never heard your meetings.
The concrete failure was an Italian–Spanish negotiation in which the word "Stripe" was in the shipped list. It is a payments company in one context and an ordinary noun in the other, and the engine had been told to expect the company. A term that is helpful in one language pair is noise in another.
Judge a term against every target language, not against English
This is the part that is easy to get wrong. When you add a term you are thinking in one language — usually the one it was coined in. But the list applies to the session whichever target language you are translating into, and a string that is unmistakable in one language is an ordinary, frequent word in another.
Before adding an entry, ask:
- Is this string also a common word in any language I translate into?
- Is it a common word in any language likely to be spoken in my sessions?
- Would I still want the engine to prefer it if it heard something merely similar?
- Will it actually be said aloud, or does it only appear in the slides?
If you can't answer those, leave the term out. Adding entries is cheap; the cost shows up later as strange words in unrelated sentences, and by then you have several suspects. The target languages we support are on the languages page.
Using one list across a team
The workable pattern for a team is: one person maintains the term file, and everyone else imports it. An administrator curates the vocabulary — reviewed against the languages the team actually works in — publishes the file wherever the team keeps shared documents, and each member imports it into their own app. When the vocabulary changes, republish the file and the team re-imports.
What it is not: there is no organisation-wide knowledge base that is applied automatically to every member's sessions. Nothing is pushed from an admin console to a colleague's app. A term file is a file: somebody has to hand it over, and somebody has to import it. If you need that for a rollout plan, plan for the import step rather than assuming distribution.
What this feature is not
- Not a translation memory. It does not store how a phrase was translated last time, and it does not enforce a preferred translation of a term.
- Not a do-not-translate list. You cannot mark a term as "leave this in the source language".
- Not automatic across an organisation. Each person imports the file into their own app.
- Not available on every target language. Qwen-routed targets only — the point the top of this page opens with.
- Not a guarantee. It biases recognition toward a string; it does not force one.
A routine that works
- Start with nothing. Run a normal session and read the transcript afterwards.
- Collect the words that actually came out wrong — that list is usually shorter than expected, and it is real evidence rather than a guess.
- Add those, checking each one against every language you translate into.
- For a specific call, add the participant names in the two fields before you start, rather than into the durable list.
- Re-read a transcript a week later. Remove entries that were never spoken, and any that started turning up where they did not belong.
None of this requires a particular plan — the term list is part of the app on every tier. See pricing for what does differ, and features for the rest of what the app does during a session.