The voice answered in the wrong language

What it looks like#

A caller speaks one language and gets answers in another, or the conversation drifts across languages mid-call. Sometimes the complaint comes from reading a transcript rather than from the call — which is a distinction worth making before anything is changed.

What decides the language#

voice_region.plan writes the rule into the brief before the call, and it writes a different one for each of four situations. Reading a complaint against the wrong one is the usual reason a working call looks broken:

1. The opening is the line's language. The brief opens in it, and the greeting's script is the language the call actually opens in rather than whatever the line's setting nominally says.
2. A remembered language is spoken from the first reply. It is not held back for proof; a whole sentence from the caller in another language is what overrides it.
3. A likely language is never assumed, and follows two or three words. Where the contact's city or state maps to a region, the call opens in the line's language and moves to the regional one the moment the caller uses it — a few words, mixed with English, are enough, and it stays there. Nothing is offered to the caller and no phrase is quoted at them.
4. With nothing on file it takes a whole sentence. A word, a name or a returned greeting is not enough to establish that somebody wants the call conducted in that language.
5. A pinned language is never steered. If the line pins a language, that decision wins over the evidence, deliberately.
6. One garbled line does not undo an established language. A single bad transcription is treated as noise, not as a change of mind.

Mid-call, one steering note may be added when a reply comes back in a script the caller has not used. On the default Gemini 3.1 model the session is immutable and that note cannot be delivered at all, so what the opening brief established is what the call runs on. That is why the rule above is written before the call rather than corrected during it.

Causes, most likely first#

1. The line's language is not what its callers speak. The commonest cause by a distance, and the easiest to fix.
2. Nothing on the contact says where they are. With no city or state on the record the caller is in the lowest tier, where it takes a whole sentence rather than two or three words. Adding the city they are actually in is often the whole fix.
3. The caller's evidence arrived late. Until it does the call is in the opening language, which is correct behaviour and can still feel wrong to the caller.
4. The transcript is being read as evidence of the spoken language. See above — romanised script is a transcription choice, not a language.
5. A pinned language is fighting the caller. Pinning is absolute; if it is set, no amount of caller evidence will move the call.
6. A model-limited behaviour is being blamed on configuration. A respect-word cadence rule was measured, changed nothing in most replies, and in one case produced an English sentence answered in another language. It was rolled back rather than left in the prompt.

What Connect completed#

• Held the conversation and recorded every turn, with the language handling on the record rather than implicit.
• Applied the evidence rule for the tier the caller was in: the opening language, then a switch on a whole sentence, or on two or three words where their city or state had already made that language likely.
• Added at most one steering note, rather than accumulating instructions that would cost latency on every later turn.
• Produced the summary, any lead and any follow-up from the call as normal — a language complaint does not invalidate the conversation's content.

What Connect did not complete#

• It did not guess the caller's language from where they are. A region makes a language likely; the caller's own words are what move the call, and with nothing on file that means a whole sentence.
• It did not override a pinned language, even with clear evidence — that is the meaning of pinning.
• It did not repair the opening. Once a call has opened in the wrong language, the first exchange has already happened in it.
• It did not change the line's setting for next time. That remains a decision for a person.

What you can do#

1. Establish what was actually spoken, not what the transcript's alphabet suggests.

   Result This alone resolves a good proportion of reports of this fault.

2. Check the line's language setting against who really calls the number.

   Result If most callers open in a language the line does not, the opening is wrong for every one of them.

3. Check whether the contact carries a city or state at all.

   Result With one on file the call follows two or three words of the regional language; without one it waits for a whole sentence. This is the difference between a call that follows the caller quickly and one that seems stubborn.

4. Check whether a language is pinned on the line, the purpose or the contact.

   Result A pin explains a call that refused to follow obvious caller evidence.

5. Try a test call in the caller's language, speaking a full sentence early.

   Result If the switch happens then, the mechanism is working and the opening language is the thing to change.

What an administrator can do#

• Set the line's opening language to the one most callers use, per line rather than per workspace.
• Keep the city or state on contact records. It is what moves a caller into the tier where a few regional words are enough, and it is inferred from a number only for an Indian landline — a mobile never names a region, because portability made the series meaningless.
• Use a pinned language only where a single language is genuinely required, and record why.
• Where the callers themselves mix, let them. A mixed line is not a compromise, it is the accurate description of the conversation — and for a caller whose region is on file, that mixed line is also what moves the call into their language.
• Do not spend prompt characters on cadence or address-form rules. They were measured, they changed little, and every character is paid for in first-token time on every turn.

When to escalate#

Escalate when the voice switches away from a language the caller has clearly established with several full sentences, when a pinned language is not being honoured, or when the greeting itself opens in a language no setting selects. Bring the call, the line's language setting, and whether any language is pinned anywhere in the resolution chain.

Questions#

Related#

• Tuning the voice, end to end
• The voice talked over the caller
• Phone and voice in Connect
• Findings a setting cannot fix
• The call sounded bad