What happens when a caller switches language mid-call
A caller rings a clinic in Copenhagen and starts in Danish. Halfway through she hands the phone to her partner, who speaks English. A human receptionist switches without thinking. A voice agent has to switch three things at once, and if any one of them lags behind, the call goes strange. This post explains what has to happen inside the agent, where it breaks, and how to set it up so that it does not.
Why it happens more than you think
Mixed-language calls are not an edge case in Northern Europe. Expats book dentists. Swedish tourists ring Danish hotels. A Finnish-speaking customer in Helsinki gets a Swedish greeting and answers in Finnish. A Polish tradesman calls a Danish wholesaler and starts in English because he assumes nobody will understand Polish. People also mix inside a sentence: a Danish caller reading an English street name, a Norwegian saying a product name in English.
If the agent restarts, apologises, or carries on in the wrong language, the caller either hangs up or shouts. Neither is a booking.
Three things have to change at once
A voice agent is three models in a row: one listens, one thinks, one speaks. Language lives in all three. We described the arrangement in Pipeline, speech-to-speech or hybrid?; here is what it means for a switch.
The listener turns sound into words. If it was told "expect Danish", it will try to hear Danish in an English sentence and produce nonsense. It has to either handle several languages at once or be told, quickly, that the language changed.
The thinker reads the words and writes the reply. It has to notice the switch, answer in the new language, and keep everything the caller already said: the name, the time, the reason. A model that answers in English but forgets the Danish half of the call has switched badly.
The speaker turns the reply into sound. If the voice was made for Danish only, its English will come out with a heavy accent or fall apart on certain words. A voice that can speak both is the difference between a switch that feels natural and one that feels like a different receptionist took over.
When people say "the agent switched language", they mean all three did, within one turn of the conversation.
How the agent follows
In NordTell the agent has a primary language, which is the language of its greeting and of its instructions. When you give it a second language, the listener expects both, the reply is written in the language the caller used, and the same voice reads it. The caller hears the same agent, in their own language, without a restart.
The safety nets follow too. The phrases that mean "put me through", "I want a person" and "goodbye" are recognised in 13 languages, so a caller who switches from Swedish to English and says "can you transfer me" gets a transfer, not a polite non-answer. The greeting says it is an AI in the primary language; if you serve two groups equally, you can write the greeting in both.
What stays fixed is the knowledge. If your opening hours and prices are in a Danish document, the agent will quote them and translate as it answers. That is fine for facts and risky for anything with legal weight, so give it the important documents in every language you expect.
Where it goes wrong
- One foreign word is not a switch. A Danish caller who says "Main Street" has not changed language. The agent should switch on a sentence, not a word, and the instructions can say so.
- Two people on one line. When a phone is passed over, the new speaker may be quieter, further from the microphone and in a different language, all at once. Expect the first sentence to be lost; the agent should ask them to repeat rather than guess.
- A voice that cannot do the second language. The switch works in the listener and the thinker, and then the speaker mangles it. Test the voice in both languages before launch.
- Numbers in the wrong language. "Halv ti" and "half ten" are an hour apart. The agent should read times back in the caller's current language, in full.
- A knowledge base in one language. The agent will quote it faithfully and translate on the fly. Fine for facts, risky for wording that matters.
How to set it up, and how to test it
Choose a voice that was made for both languages you expect; our guide to choosing a voice explains how to listen for it. Give the agent its primary language and a second one. Write one line in the instructions: "Answer in the language the caller is speaking. If you are not sure, ask." Then test from the browser: start in one language, switch mid-sentence, hand the call to a colleague who speaks the other, and say a time with "halv" in it. Listen to the recording, and do the same test with the two languages the other way round. Our guides to Danish and to Swedish and Norwegian list the phrases worth including.
Where NordTell fits
NordTell agents can switch language mid-call when the caller does, with safety nets in 13 languages: Danish, Norwegian, Swedish, Finnish, English, German, Dutch, French, Spanish, Italian, Portuguese, Polish and Turkish. The listening model and the voice are chosen per agent, and the browser test lets you try the switch before a number is connected. Whether your two languages sound right together is something you can hear for yourself in five minutes.