Fixing Names and Terms With a Glossary

The first time an organization listens back to its own translated broadcast, the complaint is almost never about grammar. General language comes out fine. What goes wrong is the handful of words that belong to you: the product nobody outside the company has heard of, the name of the person on stage, the ministry with a name that is also an ordinary noun, the acronym in every internal email.

Those are the words a model has no way to get right, because getting them right is not a language problem — it is a fact about your organization. A glossary is where you tell it those facts.

What an entry actually is

An entry is one term, written as it is spoken in your source language, plus how you want it rendered in each output language, plus an optional note. Three parts, and the second one is where people go wrong.

Term. The word or phrase as the speaker says it — up to 80 characters, so a phrase is fine, a paragraph is not.
A rendering per output language. Not one replacement for all languages: one per language you want it applied to. An entry reaches a language's translation only if you gave it a rendering for that language. This is the most common reason a glossary "does not work" — the entry exists, but not for the language somebody was listening to.
Note (optional). Guidance rather than a replacement, and the most underused field. Things like NEVER translate this as "unfair"or this is a person, not a place do more than another synonym would. The note travels with the renderings, so an entry still needs at least one — and a language with no rendering gets neither the word nor the note.

When a change takes effect

This is the part that is not visible in the interface, and it catches everybody once.

A translation session reads the glossary when it opens, and keeps that copy for its whole life. It does not re-read it. So:

Save before you go live. A glossary saved five minutes before the service applies to everything that starts after it. That is the easy case and the one to aim for.
Mid-broadcast, a save lands unevenly. A language somebody is already listening to keeps the old glossary until its session restarts. A language whose first listener arrives after you saved gets the new one. Both behaviours are correct, and they will look inconsistent if you do not know this.
Nothing is retroactive. Fixing a name halfway through does not fix the twenty minutes already translated, and the recording and transcript keep what was actually broadcast, mistakes included.

The practical rule: treat the glossary as preparation, not as a live control. If a name is coming out wrong during a service, write it down and fix it before the next one.

Three things that surprise people

Entries are filed under the language being spoken

A term is stored against the source language it is spoken in, and only the entries for the current broadcast's spoken language reach the translation. If your organization sometimes presents in Romanian and sometimes in English, those are two sets of entries, and a term you added under one will be silently absent from the other. The console's glossary shows a tab per spoken language for exactly this reason.

Adding a new output language means revisiting the glossary

Because a rendering is per language, the day you turn on a fifth language every existing entry is inert for it until you add a rendering for it. Put "check the glossary" next to "add the language" in your notes — see how to choose which languages to offer for the rest of that decision.

A hundred entries is a budget, not a filing cabinet

There is a ceiling of 100 entries, and it is more generous than it sounds because a good glossary is short. Every entry you add is a rule the model has to weigh; a list of eighty plausible-but-unimportant words is worse than a list of twenty that genuinely matter. If you are close to the limit, the problem is usually the list rather than the limit.

What belongs in it

Proper nouns that are also ordinary words. The riskiest category: a name that means something else gets translated instead of kept.
Product, ministry and department names. Anything internal that a stranger would not recognise as a name.
Acronyms. Especially the ones that collide with a common word in another language.
Words your organization has already agreed on. Where you have decided the right word for something and everyone uses it.
Names of people who appear regularly. The speaker, the leadership, the recurring guests.

What does not belong in it

Ordinary vocabulary. If a competent translator would get it right without being told, the model will too.
Whole sentences. A glossary steers word choice; it is not a script and it will not make the model say a fixed line.
Anything you are unsure of. A wrong rendering applied confidently to every session is worse than no entry at all.
Bible quotations, if you are a church. Those are handled separately: a quotation is rendered from a published Bible translation chosen per language rather than re-translated from the wording on stage. See live translation for churches.

How to find your terms without guessing

Do not sit down and try to write the list from memory — you will produce ordinary words and miss the ones that actually break. Instead, ask somebody who speaks one of your output languages to listen to two minutes of a real session and tell you what jarred. Their listening is also what creates the transcript for that language — nothing is recorded for a language nobody tuned into — so afterwards you can read it back and mark the words that came out wrong. Ten minutes of that produces a better list than an hour of imagining.

Then do it once more after the next session. Two passes get most organizations to a glossary they stop thinking about.

The short version

Fix names, not grammar. Grammar is already fine.
A rendering is per output language — an entry with no rendering for a language does nothing there.
Entries are filed under the language being spoken.
Save before you go live; a session keeps the glossary it opened with.
Twenty good entries beat eighty plausible ones.

For how the translation path works around all of this, see how it works. For settings that matter in a workplace or a classroom specifically, see live translation for business and live translation for schools.