# Norwegian Transcription on Mac: Audio and Meetings

> Plain-text version of https://kviskr.no/en/norwegian-transcription-mac (last updated 11 October 2026).
> Norwegian version: https://kviskr.no/transkribere-lydfiler

Kviskr is a Mac app that transcribes Norwegian audio files and meetings locally,
with NB-Whisper, the speech model from the National Library of Norway. You drag
in an interview, a lecture or a voice memo, or record a meeting, and get the text
with speakers and a timestamp for each line. The speech recognition and the
speaker separation run on your Mac, so the audio is never uploaded, and it works
offline. You choose bokmål or nynorsk as the written standard. The app itself is
in Norwegian only.

## How it works in Kviskr

| What | How it works |
|---|---|
| Audio formats | Files macOS can read, such as WAV, MP3, M4A, AAC, AIFF and FLAC. The original is not changed |
| Video | Not supported. Extract the audio first, for example in QuickTime Player |
| Length | No fixed limit. Long files take longer, depending on your Mac |
| Speakers | Detected locally when Kviskr hears at least two voices. Rename them as you like |
| Timestamps | Each line has a timestamp in the app. Copied text has names, not timestamps |
| Written output | Bokmål or nynorsk, which you choose. English with the same models |
| Export | Copy the text, or export recordings as a plain-text file. No Word or subtitle export |
| Internet | Not needed. The standard Norwegian model and the speaker models come with the app |
| Storage | Audio and text stay on your Mac and are deleted after 30 days unless you change it |
| Requirements | A Mac with Apple silicon (M1 or later) and macOS 15.1 or later. The app is in Norwegian |

## From recording to text

1. **Drag in the audio file.** Drop the file on the Kviskr window, click
   Importer lydfil (Import audio file) under Opptak (Recordings), or right-click
   the file in Finder and choose Open With → Kviskr. You can add several files
   at once.
2. **Let the Mac do the work.** Files wait in a queue and are transcribed one at
   a time. NB-Whisper turns the Norwegian speech into text on your Mac, and
   Kviskr finds the speakers in the same pass. The file is not uploaded
   anywhere.
3. **Show speakers, name them, copy.** Turn on Vis talere (Show speakers) to see
   each line with a speaker and a timestamp. Rename Taler 1 and Taler 2 to real
   names or pseudonyms, then copy the text or export it as a plain-text file.

## Norwegian meeting transcription

From Kviskr 0.7.0 you can record a meeting directly. Press Start møte (Start
meeting) under Opptak or in the menu bar, and press Stopp when you are done.
Kviskr transcribes the meeting and separates the speakers on your Mac. No bot
joins the call, and nothing is sent to a meeting service.

- Meetings in a room are recorded with the microphone.
- For video calls there is a setting, Ta med lyd fra Macen (Include audio from
  the Mac), that also records what the Mac plays. It is new in 0.7.0, so make a
  short test recording with your own headset and call app before an important
  meeting.
- Voices are not remembered between meetings, so you name the speakers each
  time.
- Kviskr does not write the summary. You get the full transcript.

Tell the others that you are recording. Kviskr shows that reminder the first
time you start a meeting. More, in Norwegian: https://kviskr.no/transkribere-mote

## How accurate is it on Norwegian?

NB-Whisper was trained by the National Library of Norway on around 66,000 hours
of Norwegian speech. Its paper (https://arxiv.org/abs/2402.01917) reports a 2.2%
word error rate for NB-Whisper Large on read bokmål in the NST dataset, against
6.8% for OpenAI's Whisper Large-v3. Read speech is the easy case. Free
conversation is harder, and in our own measurements most of the difference is
words that get left out when people talk fast, quietly or over each other.

Kviskr has measured its own models on public Norwegian test sets. The table shows
how many words out of 100 were left out. Standard is the model that comes with
the app. Ekko (beta) is Kviskr's own model, trained further from NB-Whisper
Large, and can be downloaded from 0.7.0.

| Test set | Words left out, Standard | Words left out, Ekko (beta) |
|---|---|---|
| NB Samtale: free conversation (500 clips) | 21.2 | 10.2 |
| NB Tale: spontaneous speech, all regions (302 clips) | 4.5 | 1.8 |
| NorGovPCC: government press conferences (200 clips) | 17.6 | 5.0 |

Per 100 words in the reference, lower is better. Ekko gets slightly more words
wrong in free conversation (8.9 against 7.4 per 100), because it writes down
words Standard skips. Measured 11 October 2026. Full method and all five test
sets, in Norwegian: https://kviskr.no/metode#egne-malinger

## Bokmål, nynorsk and dialects

Norwegian has two written standards, and the speech model cannot tell from the
audio which one you want. You choose Norsk (bokmål) or Norsk (nynorsk) in the
settings or the menu bar, and Kviskr writes in that standard whatever dialect is
spoken. Nynorsk in free conversation is the hardest case we have measured: on the
124 nynorsk clips in NB Samtale, Standard left out 24.3 of 100 words and got 10.3
wrong, against 21.2 and 7.4 for the whole test set. Names and technical terms
that keep coming out wrong can go in the vocabulary (Ordbok).

## Other ways to transcribe Norwegian

Several good services transcribe Norwegian, and some use the same speech model as
Kviskr. The main difference is where the audio goes and who may use the service.
At the University of Oslo, Autotekst runs on the university's own servers. At
NTNU, Speech to text is for public and internal data only. Jojo from the
newspaper VG is local like Kviskr. transkribere.no and TurboScribe are cloud
services you upload to.

| Tool | Where the audio is processed | Who can use it | Speech model |
|---|---|---|---|
| Kviskr | On your Mac | Anyone with an Apple silicon Mac | NB-Whisper, Ekko (beta) |
| Autotekst (University of Oslo) | UiO servers, approved up to red (confidential) data | UiO staff and students; other non-commercial organisations can buy access | NB-Whisper or OpenAI Whisper |
| Speech to text (NTNU) | NTNU service, public and internal data only | NTNU users, on the NTNU network or VPN | NB-Whisper Large or OpenAI Whisper |
| Jojo (VG) | On your Mac | Anyone with a Mac (macOS 13+) | Whisper; NB-Whisper can be downloaded |
| transkribere.no | Servers in the EU, files deleted within 24 hours | Anyone, from NOK 0.49 per minute | Not stated |
| TurboScribe | Cloud | Anyone; 3 files of up to 30 minutes a day without paying | Whisper large-v2 |

Checked on each service's own website on 11 October 2026. Prices and rules
change, so check the service before you decide.

- Autotekst: https://www.uio.no/tjenester/it/lyd-video/autotekst/
- NTNU Speech to text: https://i.ntnu.no/wiki/-/wiki/English/Speech+to+text
- Jojo: https://apps.apple.com/no/app/jojo-transcribe/id1659864300
- transkribere.no: https://transkribere.no/
- TurboScribe: https://turboscribe.ai/pricing

## Privacy

The audio and the speech recognition stay on your Mac, and there is no account.
Kviskr keeps a copy of each imported file next to the text and deletes both after
30 days unless you change it. Kviskr does not encrypt its history itself, so turn
on FileVault. The optional AI modes are off by default. When one is on, the
transcribed text, never the audio, is sent to the provider you chose.

## What you need

A Mac with Apple silicon (M1 or later) and macOS 15.1 or later. The app takes
about 1.3 GB with the standard Norwegian model. Menus and help are in Norwegian,
and you are welcome to email hei@kviskr.no in English. For dictation into any
app, see https://kviskr.no/en/norwegian-dictation-mac

## Frequently asked questions

### Can I transcribe Norwegian audio offline on a Mac?

Yes. Kviskr runs NB-Whisper locally, and the standard Norwegian model and the
models that separate speakers come with the app. Once it is installed, you can
transcribe audio files and meetings without an internet connection.

### Does it handle nynorsk and Norwegian dialects?

You choose bokmål or nynorsk as the written standard, and Kviskr writes in that
standard whatever dialect is spoken. It cannot tell the standard from the audio.
How well it works varies: free conversation, dialects that are rare in the
training data and distant microphones give more errors.

### Can Kviskr tell who is speaking?

Yes. Kviskr separates the voices locally and labels the lines Taler 1, Taler 2
and so on, each with a timestamp. You rename them afterwards. Voices are not
remembered between recordings.

### Can I get the transcript in English?

Kviskr transcribes what is said. For a translation, you can turn on one of the
optional AI modes, which sends the transcribed text to Anthropic or OpenAI. The
audio is never sent. The AI modes are off by default.

### Does Kviskr write meeting summaries?

Not on its own. Kviskr gives you the full transcript with speakers and timestamps. A summary
needs a large language model in the cloud, so you either paste the text into an
AI tool you trust or turn on an AI mode, and in both cases the text leaves your
Mac.

### Is the app available in English?

No. The interface and the help pages are in Norwegian only. This page uses the
Norwegian names of buttons, with an English translation in brackets.

### Does Kviskr run on Windows or iPhone?

No. Kviskr is a Mac app and needs Apple silicon (M1 or later) and macOS 15.1 or
later. There is no Windows or iPhone version, and no date has been promised.

Download: https://kviskr.no/last-ned