Make language switching explicit in a voice conversation

Choose a preference the caller can correct

Start from an explicit authenticated language preference or the conversation’s requested purpose. A telephone country code is routing information, not proof of the caller’s preferred language, nationality or demographic identity. If the preference is unclear, ask a short neutral clarification that the caller can answer. Keep task facts separate from delivery preferences so switching from English to French does not change the appointment date or intended recipient. A language setting should guide recognition and spoken phrasing while preserving the same workspace and resource boundaries.

Use recognition signals cautiously

A short utterance can contain an English product name inside a French sentence or an acknowledgement common to several languages. Automatic detection is useful evidence, but a single label should not override a clear user preference. Lupin’s cascade requests language detection from Scribe and prepares recognition per utterance, while native audio has different managed behavior. Test both under their actual captured mode. When a language transition is ambiguous, confirm it briefly rather than repeatedly switching back and forth and making the conversation harder to follow.

Validate the voice and model combination

Provider language support is a property of the chosen model and voice configuration, not a universal promise attached to the application. The current cascade uses V4 Turbo synthesis; other modes can report managed components. Check provider metadata and a short synthetic listening sample in the target language before declaring support. Include accented names, dates and an identifier that must remain unchanged. Do not select a voice by inferred gender or persona. Evaluate intelligibility, pronunciation and reliable switching for the actual task the owner wants to complete.

Keep approval text anchored to exact arguments

A translation can change the apparent meaning of an action if it paraphrases a recipient, amount or account. Show the exact structured action in the authenticated approval surface even when the spoken explanation uses another language. Lupin’s call conversation cannot execute a sensitive action simply because the caller says an approval phrase. For example, a French explanation of an English calendar title should preserve the original title and time zone in the pending request. Rediscover and reapprove if the account revision or tool arguments change during the conversation.

Test transitions rather than isolated sentences

Create a dialogue that starts in one language, requests a switch, returns a tool result containing names in another language and then corrects a date. Assess whether the system retains the chosen response language and the corrected task facts. Include a brief code-switch that should not change the overall preference. Record recognition errors separately from synthesis pronunciation errors so the right component can be adjusted. A useful result is a repeatable set of synthetic transitions with explicit expected states, not a claim that every supported provider language has equal quality.

Keep a mixed-language task example

A caller can request French replies while preserving the literal English calendar title Product review. The system should speak French around that title, keep the canonical title unchanged and display the exact date and time zone in the approval. This tests language delivery without accidentally translating the resource the action will create.

Sources & verification

Reviewed against current checked-in Lupin routes, shared contracts and official provider references. No live calls, OAuth authorizations, app actions or purchases were performed.

• Lupin phone, SMS and calls ↗Checked 2026-11-05
• ElevenLabs realtime speech recognition ↗Checked 2026-11-05
• ElevenLabs model reference ↗Checked 2026-11-05

Prepared with AI assistance by Lupin editorial. How we research and maintain these guides.