A receptionist can greet a caller in two languages and still fail when the caller gives an address, changes language halfway through, or reads out a reference number. Language coverage has to reach the saved record.
For a multilingual agent, I would define supported tasks in each language before writing a claim about how many languages it handles.
What does multilingual support actually include?
The speech recognizer, conversation model, voice output and connected systems all contribute. A supported language in one component does not guarantee that the complete workflow works in that language.
Even the same task can need different handling. Names may be spelled using a different alphabet. Dates can have different spoken forms. A customer system may require a normalized phone number while staff need the caller's original words for context.
Keep the original confirmed value when it matters. Translating a person's name or silently rewriting a service address can destroy the information the office needs.
Write a language-and-task test matrix
Start with the languages you expect actual callers to use. Ask a competent speaker to review the scenarios and resulting conversation. A model grading its own transcript is a limited check of whether the audio was understandable.
This is an illustrative test set for an intake workflow:
| Scenario | Inspect after the call |
|---|---|
| Caller gives a local name | Confirmed spelling and saved name |
| Caller switches language | Correct next question without losing collected details |
| Caller corrects a spoken number | Final normalized value and confirmation |
| Caller uses an unfamiliar place name | Clarification or approved fallback |
| Agent cannot understand after clarification | Staff handoff instead of a guessed field |
Keep separate results for language, task and audio conditions. Passing a quiet test in one language is not evidence for a noisy call in another.
Normalize fields after understanding them
A multilingual workflow benefits from a clear separation between what the caller said and what the backend requires. Here is a proposed LiveKit and Next.js example:
Caller -> LiveKit agent -> Confirmed conversation fields
|
v
Next.js intake validation
|
v
PostgreSQL requestThis is a proposed application boundary, not a claim about a specific client deployment. Choose speech and model services after checking their current language support for your deployment.
For dates and contact windows, clarify the timezone or location when ambiguity could affect the result. Avoid sending free-form phrases directly into scheduling logic that assumes one locale.
A useful rule is to ask for clarification when interpretation changes the action. Repeating a sentence with greater confidence does not resolve an ambiguous date.
Test language changes at the integration boundary
During my focused Lawtte AI engagement, I worked on prompts, tool calls and conversation tuning for an intake product with multilingual capabilities. That is relevant experience, but it is not a published benchmark of every supported language.
The lesson I would carry into a new build is to test the full action in each agreed language. Use fictional enquiries and inspect the backend record. For a business with sensitive intake, have its team approve the task boundaries, data handling and escalation wording before launch.
Also check the staff experience. A translated summary can be useful, but the original confirmed details may still be necessary for follow-up. Specify which language staff see and who reviews uncertain cases.
Questions about multilingual receptionists
Can one agent handle callers changing languages?
Some configurations can, but test the exact language pairs and transitions you intend to support. Do not infer that capability from a translated opening line.
Should we launch every available language at once?
I would start with the languages the team can test and support. A smaller, verified scope is easier to maintain than an unsupported language list.
What if we cannot assess the quality ourselves?
Bring in a competent reviewer for the language and task. Automated transcripts can miss pronunciation or interpretation problems that are obvious to a caller.
Use the LiveKit production-readiness guide for broader engineering considerations, then add your own language-specific acceptance examples before scoping a build.