How to write role context for live translation

The same source word can need a different translation in skincare, apparel, and electronics. A brand-name list alone cannot settle every ambiguity; the translator also needs to know who is speaking, to whom, and for what task.

ObsTrans AI Roles manage both layers: role context describes the situation, while a glossary records source → preferred target pairs. In the product implementation, speech translation injects active role context and terminology constraints into supported translation routes; the real-time path also passes active speech glossary pairs into its translation configuration. Live-chat translation is a separate scene, so do not assume the speech role automatically governs comments.

What useful role context contains

Answer four questions:

1. Who is speaking? Brand representative, factory demonstrator, or reviewer.
2. Who is listening? Category newcomer or specification-aware buyer.
3. What is happening? Demonstration, after-sales Q&A, or campaign explanation.
4. What is forbidden? Added medical effects, guarantees, or facts absent from the source.

For example:

Translate a skincare live stream for general US consumers. The host is a brand representative explaining application order and ingredients. Keep the language conversational and concise, preserve model names, and never turn personal experience into a medical claim.

That provides more signal than “translate professionally” and less noise than a complete catalogue.

What belongs in the glossary

Brands, models, ingredients, series names, and campaign names require exact output. Google’s official Cloud Translation glossary guide defines a glossary as a way to translate domain-specific terminology consistently, typically including named entities.

Context can say “this is skincare.” It cannot guarantee identical spelling and capitalisation for a brand on every occurrence. Exact renderings should not be buried in prose.

The current ObsTrans role editor accepts up to 50 term pairs. That verified product limit does not mean filling all 50 improves quality. Our two-layer glossary guide explains the upstream boundary: a translation glossary cannot match a brand name already misheard by speech recognition.

Keep unstable facts out

Do not put these in persistent context:

• changing prices and inventory;
• unconfirmed shipping promises;
• the full verbatim show script;
• instructions to embellish or add selling points;
• conflicting markets or language directions.

Role context should survive several broadcasts. Temporary facts must be spoken, preferably with the short and explicit structure in the script localisation guide.

Accept it with a difference table

Choose five minutes of real audio containing an ambiguous category word, three names or models, one negation, and one promotional condition. Run with the role off and on:

Source	Without role	With role	Improvement?	Regression?

Counting corrected target terms is not enough. Context may fix a category word while making ordinary sentences too promotional. Every change needs a regression check against output that was already correct.

Split by category, not presenter

One presenter can sell skincare today and headphones tomorrow; the context changes completely. Two presenters selling the same category to the same market share much more terminology. Scope roles as category + target market + task.

ObsTrans lets users copy a preset into their own role, edit it, and make it active. A preset is a starting template, not a live guarantee. Accept it on your products.

What context cannot replace

Role context cannot repair microphone capture, endpointing, or deleted words, and it cannot determine whether a marketing claim is lawful. It acts at the text-translation stage. Use the speech recognition diagnosis first to prove the input text, then measure what context changes.