One Glossary, Nine Languages: Why We Didn't Copy the Dictionary-per-Pair Model

TL;DR — A glossary model where every language pair needs its own dictionary makes your term list grow faster than your market: 50 brand terms × 7 target languages = 350 entries to maintain. We made the opposite bet: one table, nine languages, per-line proof. Here's why.

The pattern everyone copied from machine translation

Open a glossary feature in most AI video tools today and you'll meet the same shape: "Create dictionary → pick source language → pick target language → add pairs." One dictionary governs exactly one language pair. It's inherited from classical statistical MT, where phrase tables genuinely were pair-scoped — but it forces a human cost the architecture no longer deserves:

Per-pair dictionariesVoxtakes term rules
Entries for 50 terms → 7 languages350 — or lazily, only the pairs you happened to fill50 + overrides only where a language genuinely differs
New language added next quarterRe-enter your whole list for that pairZero — existing rules already govern it
"Never translate this" (brand verbatim)Repeat it in every dictionaryOne keep verbatim flag, all languages
Did the dub actually obey?Spot-check by eyePer-line ✓ locked / ⚠ slipped badges on the review line itself

What a term rule actually is here

Each entry is one record: the source term, an optional global default rendering, an optional per-language override map, or a keep-verbatim flag. When a job renders, the translator is constrained by your current table — and every finished line is machine re-checked against it. A green ✓ badge on the review page means "this line demonstrably locked your term"; an amber ⚠ means it slipped, and you fix that one line instead of re-rendering the video and hoping.

Why the proof badge matters more than the table

A glossary that "probably" steers the model is a placebo with an import button. The enforcement loop is the product: constrain → render → check every line → surface misses where you'll actually see them. Protection isn't a policy you trust — it's a status you read, line by line.

And when your brand evolves

Every job records the exact count of terms that guarded that render. When your library has moved since, the 📘 chip on the finished video offers a one-click re-render with current terms (billed like any job — Economy delivery still halves it). No silent staleness: just the option, where you'd look for it.

Try it with your own brand terms: open a free account — verified emails get 5 free credit minutes, no card.

← All articlesStart a free dub →