The longer you talk, the sharper it translates

Pin product names and technical terms to your exact wording, while recurring terms are auto-collected live and applied losslessly at sentence boundaries.

What it is

A glossary is a set of proper nouns and technical terms. The active set is injected automatically at session start, so both speech recognition and translation prefer your exact wording.

Smart context auto-collects terms that keep coming up in live speech and proposes them as candidates. Press [Apply now] and they take effect at the next sentence boundary without losing a single caption — a loop that gets more accurate the longer the talk runs.

When to use it

Product and company names keep getting translated differently, eroding trust
Technical terms must go out verbatim in English, or in your team's standard translation
You want to reduce misrecognition by describing the topic and background up front
A session started unprepared and terms need refining mid-live

How to use it

  1. [Create new set]

    Create a new set in the console's Glossary. First time? Use [Insert sample set (IT conference)] to see the format.

  2. One entry per line

    The source=translation form forces a translation; without =, the entry boosts recognition. Lines starting with # are comments.

  3. Write the background context

    Describe the topic and product background in free text — it's injected alongside the entries as a recognition and translation hint.

  4. Check [Use as active]

    The active set is injected once, automatically, whenever you start a live caption session.

  5. [Apply now] while live

    Add the candidates smart context has collected, press [Apply now], and they take effect losslessly at the next sentence boundary.

The format — one source=translation per line

A glossary set is plain text. One entry per line — source=translation forces a translation, a bare term boosts recognition, and lines starting with # are comments.

# IT conference sample — one entry per line, source=translation
Newey
Kubernetes
LLM=large language model
RAG=retrieval-augmented generation
fine-tuning
GPU cluster

Registered terms apply automatically at session start, and both speech recognition and translation prefer your exact wording.

Expert glossaries — presets ready to switch on

Packs curated by Newey for the fields where jargon trips up translation — legal, finance, medical, and more. Switch one on and pick it for a session, just like a specialized set.

Legal

Contracts, litigation, and courtroom terminology.

45 terms

Web3

Blockchain, crypto assets, and on-chain jargon.

48 terms

Finance

Markets, investing, and corporate finance terms.

49 terms

Programming

Software engineering and developer jargon.

56 terms

Japanese anime

Anime and manga culture, including Japanese loanwords.

45 terms

Slang

Internet slang and casual expressions.

46 terms

Gaming

Game mechanics, esports, and community jargon.

62 terms

Medical

Clinical, diagnostic, and healthcare terminology.

63 terms

AI & machine learning

Model names, labs, and ML terminology.

60 terms

Startup & business

Metrics, fundraising, and workplace jargon.

60 terms

K-pop & K-content

Fandom culture, Korean loanwords, and entertainment terms.

58 terms

Packs are provided read-only — duplicate one to turn it into a set of your own. They carry recognition-boost terms only, so they work with any language pair.

Smart context — a glossary that learns as it listens

As live speech accumulates, recurring-term candidates refresh every 30 seconds. AI suggestions bundle terms, translation pairs, and background context — only the items you select are added to the glossary.

Applies losslessly at the next sentence boundary

Auto-collection

Terms that come up often in the talk are proposed as candidates.

[Apply now], losslessly

New terms take effect at the next sentence boundary without a caption gap. Even if you never press it, they apply automatically from the next reconnect or session rotation.

Longer means sharper

In an hour-long keynote, the second half translates more accurately than the first — the terms have been piling up.

Requirements & limitations

  • The active set is injected once at session start — glossary edits made while live apply from the next session (smart context's [Apply now] is the exception).
  • Glossaries are stored encrypted in this browser with AES-256-GCM — they stay on your device and never go to a server.

Whatever you watch, whoever you talk to

Whoever you talk to, wherever you are, connect with people everywhere in your own language.