The short answer
The right tool depends on how sensitive your data is and whether you also want to analyse it in the same place.
- Sensitive interview data at a Swedish university: check whether your university offers Sunet Scribe. Lund, Karlstad and Mid Sweden University have approved it for sensitive personal data. It runs in Sunet's data centres in Sweden and deletes files after seven days.
- Free and fully offline: run KB-Whisper from the National Library of Sweden on your own computer. It has the best published accuracy for Swedish, but does not separate speakers.
- Transcription and analysis in one place: tools like Kahubi transcribe in the EU and let you code themes with every quote linked to the recording.
- Avoid Otter.ai for Swedish: it does not support the language, and its policy allows training on your recordings.
Swedish transcription tools compared
For research interviews, three things matter as much as accuracy: where the audio is processed, whether the provider trains on it, and whether you can tell speakers apart. "USA" means personal data leaves the EU, which usually needs extra safeguards and your university's approval.
| Tool | Swedish | Where data is processed | Speakers | Trains on your data | Price |
|---|---|---|---|---|---|
| Sunet Scribe | Yes | Sunet, Sweden | Yes | Not stated | Through your university |
| KB-Whisper (local) | Swedish only | Your own computer | No | Not applicable | Free (Apache 2.0) |
| Kahubi | Yes | EU (France, Germany) | Yes | No | 30 min free; $29/month incl. 10 h |
| Amberscript | Yes | Frankfurt, Germany | Yes | Third parties: no | From €19/month |
| Happy Scribe | Yes | EU, some subprocessors outside | Yes | Only if you opt in | From $17/month |
| Word Transcribe | Yes | Microsoft 365 (OneDrive) | Yes | No | In Microsoft 365, 300 min/month |
| NVivo Transcription | Yes | Not stated | Yes | Not stated | 15 min free, then paid |
| Sonix | Yes | USA | Yes | No | $10/hour or from $25/month |
| TurboScribe | Yes | USA | Yes | No | Free 3 files/day; $20/month |
| Otter.ai | No | USA | Yes | Yes, de-identified | Free 300 min/month |
Kahubi publishes this guide, so its row is highlighted. Prices are list prices before VAT.
Test it on your own interview
Upload 30 minutes of Swedish audio and compare the transcript with your current tool. Processed in the EU, never used for training.
How accurate is automatic Swedish transcription?
Word error rate (WER) is the share of words a model gets wrong; lower is better. These are published results on the FLEURS Swedish test set, which is read speech, so expect more errors on real interviews with overlapping talk and dialects. Commercial services rarely publish Swedish error rates, so test any tool on 10 minutes of your own audio first.
| Model | WER on FLEURS Swedish | Source |
|---|---|---|
| KB-Whisper large | 5.4% | KBLab model card |
| KB-Whisper small | 7.3% | KBLab model card |
| OpenAI Whisper large-v3 | 7.8% | KBLab model card |
| OpenAI Whisper large-v2 | 8.5% | Whisper paper, table 13 |
| OpenAI Whisper small | 20.6% | KBLab model card |
GDPR checklist for interview recordings
When in doubt, ask your university's data protection officer. Many Swedish universities publish which tools are approved.
- Your consent form says the recording will be transcribed, and by what kind of service.
- The service processes data in the EU/EEA, or your university has approved the transfer.
- There is a data processing agreement (DPA) with the provider, usually signed by the university.
- The provider does not train its models on your recordings.
- You know how long the provider keeps files, and delete them once the transcript is checked.
- Names and identifying details are replaced before analysis or sharing.
- Sensitive personal data (health, ethnicity, religion, politics, sexuality) is covered by your ethics approval.
Getting a good transcript
Most transcription errors start with the recording, not the software.
- Record with a microphone close to each speaker. Avoid cafés and echoing rooms.
- Name the speakers (Interviewer, Participant 1) as soon as the transcript is ready.
- Listen through every passage you will quote, and spot-check 10% of the rest. Dialects, names and numbers are the usual errors.
- Pseudonymise names, places and workplaces before coding.
- Keep each code linked to its timestamp so you can return to the voice, not only the text.
Last updated 2026-10-07.