Voice AI in Swedish and Norwegian, without the accent
Swedish and Norwegian are often sold as one box on a language list, next to Danish. They are close cousins, and a Swede and a Norwegian can chat without switching to English. But a voice agent does not get to nod along and guess. It has to hear a Bergen dialect, say Göteborg the way a Swede says it, and tell two words apart that differ only in melody. This post is about where voice AI stumbles in the two languages, and how to find out before your callers do.
Norwegian: the written language is not the spoken one
Norway has two written standards, Bokmål and Nynorsk, and no standard spoken language at all. People speak their dialect at work, on the radio and on the phone, and nobody is expected to change it. A caller from Bergen says "ikkje" where the written form is "ikke"; a caller from Trøndelag says "æ" for "jeg"; a caller from the north may say "ka" for "hva". The words on the page are the same, the sounds are not.
Listening models are usually strongest on the speech of Oslo and its surroundings, because written Bokmål sits closest to it, and weaker the further the caller is from the capital. There is no way round this except to test with the callers you actually have. If your customers are in Stavanger, a demo with an Oslo caller tells you very little.
Numbers add a small trap of their own. Norway changed its counting system in the 1950s, so "tjueen" and "en og tyve" both mean 21 and both are in daily use, sometimes from the same caller. The agent has to accept either and read back one.
Swedish: the melody carries meaning
Swedish has a pitch accent. Two words can be spelled the same and mean different things depending on the tune they are said to. "Anden" is the duck with one melody and the spirit with another; "tomten" is the plot of land or the Christmas gnome. Context usually saves a human listener. A model that ignores pitch loses one clue it needs, and the errors pile up on names and short words.
Then there are the regions, again. The Scanian r in the south, the Gothenburg lilt, Finland-Swedish with its flatter melody and different rhythm. A voice agent that answers for a business in Malmö will hear Scanian all day, and a business in Helsinki will hear Swedish that many Swedes find unusual. Both are correct Swedish. Neither is what the demo was tuned for.
Times work like Danish: "halv tio" is half past nine, and "fem över halv tio" is 9:35, which no English-trained model expects. Personal identity numbers come up in some businesses; they are sensitive, and the agent should ask for them only when the task really needs it.
Where both go wrong on the phone
- Place names. Göteborg, Ålesund, Tromsø, Bodø, Växjö. A voice that reads them with English vowels loses the caller's trust in a word. Listen to the voice read your ten most common towns before you choose it.
- Names with æ, ø, å, ä and ö. The listener must hear them, the thinker must keep them, the speaker must say them. A name that comes back as "Solveig" when the caller said "Sølvi" is a small thing that feels big.
- Spelling out. Norwegians and Swedes spell names letter by letter, sometimes with a word for each letter. The agent should read the spelling back.
- The accent. Even with perfect hearing, a voice built for English and asked to speak Swedish sounds like a friendly tourist. Callers hear it in the first sentence. The voice has to be made for the language.
- Danish, Swedish and Norwegian in one call. A Copenhagen business hears all three; a Malmö business hears two. The agent should follow the caller's language rather than insist on its own. See what happens when a caller switches language mid-call.
Test it before it answers
Write ten lines per language and have a native speaker say them to the agent from a mobile phone, outdoors, not from a laptop in a quiet office. Include a phone number, an address, a time with "halv" in it, a date, a name with a special letter, a place name, a request to be put through ("kan du sette meg over til" or "kan du koppla mig till"), a question the agent cannot know, and a goodbye. Then have someone from a different region do the same ten. Count the misses, look at where they cluster, and fix the listening model or the voice before you touch the instructions. Our guide to testing a voice agent before launch has the full routine, and our post on Danish lists the traps that language adds.
Where NordTell fits
Norwegian and Swedish are two of the 13 languages with built-in safety nets in NordTell, alongside Danish and Finnish. The listening model and the voice are chosen per agent, the provider is printed next to each, and you can call the agent from the browser and hear it before any number is connected. We cannot promise every dialect; we can promise that you will find out in a test rather than on a customer.