Multilingual terminology governance for voice AI

Language quality is often judged by fluency. In a real customer operation, fluency is only the beginning. A voice agent can sound natural while mispronouncing a product, translating a policy too broadly or using a local term that means something different to the customer. These errors are expensive because they change what the caller believes will happen next.

Terminology governance gives the operation a durable way to manage meaning. It connects approved names, spoken forms, definitions, language variants, policy implications and workflow owners. Dring's multilingual platform has a 62-language technical capability inventory across voice, WhatsApp, SMS and email. The public launch-priority set is ten languages, and the quality of each requested locale/workflow still depends on a maintained operating vocabulary plus actual-path validation before production.

Why a glossary is not enough

A glossary normally stores a preferred written term and its translation. Voice AI needs more. It needs the way a caller may say the term, the way the agent should say it, common misspellings, nearby concepts, forbidden substitutions and the workflow where the term matters. It may also need a policy note: whether the word signals a refund, an appointment, a financial product or a clinical boundary.

Build terminology records as small operational objects. Include the canonical term, spoken variants, languages, definition, examples, source owner, last review date and test scenarios. Keep brand names and product tiers separate from generic concepts so a model does not treat them as interchangeable.

Assign ownership across teams

Product should own product names and current descriptions. Legal or compliance should review regulated or sensitive language. Local operators should validate natural usage. Support should identify caller phrasing. Quality should turn disagreements into test cases. Without ownership, a glossary becomes a document everyone trusts until it is wrong.

Use a change process for high-impact terms. A new plan tier, refund rule or treatment name should not reach production only through a prompt edit. Record the change, affected languages, impacted workflows and release version. Dring's Agent Factory can use that information to generate targeted simulations instead of re-testing the entire operation blindly.

Govern spoken and written forms together

Customers may read a term in a WhatsApp message and then say it differently on a call. The agent should connect both forms to the same concept. If a product code is written with digits but spoken as words, normalise the concept without losing the original. If a local expression maps to a standard policy category, store the structured category while preserving the customer's wording for context.

Consistency does not mean literal translation. A formal phrase in one language may sound distant or unnatural in another. Ask local reviewers whether the agent's version is clear, respectful and appropriate for the channel. The outcome and policy must remain stable even when expression changes.

Use terminology to improve recognition

Rare names and domain phrases are often difficult for general speech recognition. Speech adaptation techniques can increase the likelihood that a recogniser chooses a known phrase, particularly for proper names, unusual words and noisy audio. Google's Speech-to-Text documentation describes phrase sets and custom classes as ways to bias recognition toward domain vocabulary.

Do not add every word to every language model. Use the workflow map to prioritise terms that affect identity, routing, action or compliance. Review false positives as well as missed words. A recogniser that hears a similar product name too often may create a different class of error.

Test meaning, not just spelling

For each important term, test a caller saying the canonical form, a natural local variant, a correction, a noisy version and an ambiguous near-match. Check the transcript, intent, tool parameters, spoken response and CRM field. A correct transcript that leads to the wrong action is still a failure.

Review terminology by language and workflow. A phrase can be harmless in a status call and risky in a payment call. Keep an exception list for terms that require a human when the recogniser or context is uncertain. Dring's quality controls provide a place to track these cases over time.

Make updates part of monthly operations

Products, prices, routes and policies change. A terminology review should be part of the operating calendar, not a response to a public mistake. Each month, review newly observed caller terms, unresolved language mismatches, pronunciation feedback and changes from product or compliance teams. Retire obsolete terms and confirm the test suite still reflects current offerings.

The review can also inform Sector Insight. Recurring words may signal a product opportunity, price sensitivity, service friction or a new market need. Terminology is not only a quality asset. It is a view into what customers are trying to say.

Turn glossary changes into release notes

A terminology change should leave a small, readable record. Note the old form, the approved replacement, the languages affected, the reason for the change and the workflows that need to be tested. If the change affects a product name, price or policy phrase, include the owner who confirmed the customer-facing meaning. This gives support, product and quality teams one explanation to work from when a caller uses an older expression.

Release the change in a controlled slice when the term is important or ambiguous. Compare recognition, pronunciation, intent selection and downstream records before expanding it. Keep the previous version available for rollback, and record examples from real conversations without carrying unnecessary personal details into the test set. A clear release note makes the 62-language technical inventory easier to govern across channels; ten languages are public launch priorities, and each requested locale/workflow needs actual-path validation before production. That evidence is easier to improve through the Agent Factory.

A governance pattern for 62 languages

• Maintain one concept record with language-specific expression and pronunciation fields.
• Assign owners for product, policy, local usage and quality review.
• Connect terms to the workflows where they affect routing or action.
• Test canonical, variant, correction, noise and ambiguous forms.
• Track transcript, intent, tool and CRM outcomes together.
• Review new and retired terms on a recurring release cadence.

Multilingual terminology is the quiet infrastructure behind a natural conversation. When it is governed well, customers hear familiar language, teams see reliable records and the agent improves without losing the promise that made the workflow trustworthy.

A useful test is to ask whether a new teammate could understand the term record without opening the prompt. If the answer is no, the knowledge is still too implicit. Make the decision visible: what the phrase means, when it applies, how it is spoken, who approved it and what happens when confidence is low. That small discipline helps every language move with the operation instead of drifting away from it.

Further reading

• Google Cloud on speech model adaptation, phrase sets and custom vocabulary
• Google Search Central on language versions and localization
• NIST on measuring and documenting AI behaviour across the lifecycle