Threll.ai
Dialect

Your voice agent understands the dialect. It doesn't answer in one.

Speech technology is built to turn dialect into standard writing. That is the right call on the way in — and it leaves a decision on the way out that almost nobody has made.

2 min read
Wide view of a hotel reception area in low morning light, where a guest in a dark coat with a suitcase stands at a long oak counter talking to a receptionist in a dark blue shirt, while another member of staff further back on the left holds a black telephone handset to her ear

“Does it understand dialects?” is usually the first question a business asks about a voice agent. It is a good question, the answer has got noticeably better, and it still only covers one direction.

NB-Whisper, the National Library of Norway's Norwegian speech model, is trained on material including the NST corpus and recordings from the Norwegian parliament, and it returns normalised Bokmål or Nynorsk — whoever is speaking. That is not a weakness. It is the whole point: a transcript should be readable and searchable, not a rendering of how someone said it. The dialect is heard, and then it is set aside at the first step.

The resources for the other direction exist too. The same library's pronunciation lexicon, NB Uttale, holds 785,000 words transcribed in five pronunciation variants — East Norwegian, Southwest Norwegian, West Norwegian, Trøndersk and North Norwegian — built precisely for speech technology, synthesis included. NB Tale holds recordings of 380 speakers from 24 dialect areas.

And yet the voices you actually choose between when you set an agent up are, in practice, the standard variety. One option, presented as a list.

Want to talk it through with someone who has done this before?

Fifteen minutes on a call. We will tell you honestly whether a voice agent fits your setup – and what has to be settled before it goes live.

Book Meeting (opens in a new window)+47 24 12 66 30

15 minutes · no obligation · pick your own time

It is a choice, not a setting

For most businesses the standard variety is the right answer. It is neutral, it is understood everywhere, and nobody notices it.

But it is not right for everyone. A workshop in a small coastal town, a clinic on an island, an association answering its own members: sounding local there is not decoration, it is the claim itself. A capital-city voice saying “hello, you've reached us” says something about where “us” is — without meaning to.

The seam shows most clearly at the handover. The agent speaks the standard variety for ninety seconds, and then a colleague takes over who speaks the local dialect. To the caller that is not two voices. It is one conversation that changes company halfway through. What travels across to the human is something you decide in advance — how the two sound next to each other is something few people have considered.

Two things that cost nothing

Write down why your agent sounds the way it does. Was the voice chosen, or was it first in the list? Choosing a voice is becoming a purchase with a provenance and a contract behind it, so it is worth having a reason that is not “it was there”.

And test comprehension in the direction that is actually hard. Don't read a sentence into a demo — call your own number, and have someone from each of the areas your customers call from give a surname and a date. The phone line has already thrown away a good deal of the signal before the model gets it, and that is where the differences between dialects grow, not shrink.

Threll.ai builds voice agents in Norwegian, Swedish and Danish. The dialect question has two sides, and the second — what the agent itself sounds like — has been settled at most companies without anyone deciding it.

Related Articles

Continue reading more articles

See all articles