General MT models are fluent averageers. Feed one a made-up product name and it picks whatever feels statistically right: translate it (“Zelos” → “zeal”), half-transliterate it into a form your audience can say but never search, or blend it into a similar word mid-sentence. And because the dubbing voice reads exactly that broken text, your launch video now mispronounces your own brand — in every language at once.
Why a dictionary isn't enough
Tools that “support glossaries” usually just pass them as hints; hints lose to fluency. The production pattern is two-stage:
- Constrain the translator with a locked list — verbatim keeps (brand stays exactly as written) or pinned per-language renderings (official forms, e.g. a Japanese katakana). These tokens are marked untouchable.
- Enforce afterward: machine-check every rendered line by re-transcribing the finished audio and verifying term presence. A term that slipped is flagged on that line, before you ship — a review badge, not a surprise on YouTube.
Edge cases to decide up front
- Inflecting languages: German puts articles and case endings around brands (des ZELOS?) — pin them per language once, reuse forever.
- RTL scripts: Latin names inside Arabic need directional handling or the burned caption turns to soup.
- Agencies: per-project lists so Client A's locked terms never appear in Client B's renders.
Terms are cheap to lock and expensive to get wrong on screen. Put the list in before the first render, not in the fix pass afterward.
Voxtakes glossaries lock terms across 9 languages with per-line enforcement checks — beta applications open.