Your interviews never leave.
Forty interviews at six hours of typing each is a research assistant's spring. Uploading them to a consumer transcription service is a data breach with a subscription. TRANSCRIPTION does it inside your own perimeter — speaker separation, Danish and other languages, timestamps, and names masked on the way in.
There are three ways research interviews get transcribed. A student types them. A bureau is paid per minute. Or somebody uploads them to whatever service is free this year.
The third one happens more than anybody admits at a steering group, and it is the reason this platform exists. A recording of a patient describing their illness is special-category personal data being sent to a processor nobody assessed.
The first two are defensible and slow. Six hours of typing per interview hour is the usual figure, and a bureau still means the audio travels — with a data processing agreement if you are lucky and diligent.
TRANSCRIPTION is the fourth option, and the model runs where the audio already is. It has been in production inside the Capital Region of Denmark's own datacentre, which is the best evidence we have that this works.
Forty-two minutes, in about four.
A research interview arriving as text: speakers separated, timestamps anchored to the audio, and a spoken name replaced before anybody reads it. Watch the third line closely. That substitution decides whether you can share the transcript with your team.
Two speech models. One perimeter.
TRANSCRIPTION runs on our own speech model or on NVIDIA's, and you choose which. Both run inside your perimeter, so the choice never changes where the audio goes.
- Our model
- viceron-asr, our own speech model. It has the lowest word error rate on all five Danish test sets in the benchmark below.
- NVIDIA's model
- NVIDIA-Nemotron-3.5-Transcribe-30B-A4B, NVIDIA's open-weight speech model. It transcribes 40 languages, Danish among them.
On Danish speech, viceron-asr makes 70% fewer word errors than Whisper large-v3.
Word error rate is the share of words a model gets wrong, so lower is better. All three models are scored on the same five public Danish test sets, which cover conversation, read-aloud text and speeches in parliament. Your own recordings will score differently, which is why we measure those too.
Word error rate, % · lower is better
From a folder of recordings to a pipeline.
Measure your audio
Two or three representative recordings, transcribed and human-corrected, so the accuracy conversation starts from measured numbers.
- Word error rate per condition
- Correction time estimated
- An honest go / no-go
Agree the perimeter
Where it runs, what it touches and who may see a transcript. Your IT and DPO are part of this conversation from the first week.
- Deployment target chosen
- Data flow diagram for the DPIA
- Access model agreed
Deploy and tune
Installed, connected to your storage, style and masking rules configured, and the vocabulary for your field loaded.
- Style: verbatim or cleaned
- Masking rules and key custody
- Domain vocabulary added
Run and hand over
Your team uses it themselves. We train the people who will actually be correcting transcripts, and stay reachable while the first batch goes through.
- Training for correctors
- Queue monitoring
- Named engineer for the first batch