An hour-long advisory call creates half an hour of write-up work. A cloud note-taker takes that work off your hands, and sends the recording to servers in the United States while it's at it.
Local transcription is turning speech into text with a speech model that runs on your own hardware, so the recording and the transcript never leave your network. The open models that can do this in 2026 match cloud services on Dutch benchmarks, and they run on a single workstation GPU. The difference from a cloud note-taker is no longer quality. It's where the conversation ends up.
Where does a recording go with a cloud note-taker?
Every service handles it differently, and the differences are in the privacy policy, not on the pricing page. Three examples, in the vendors' own words:
- Otter.ai says in its privacy policy (effective 16 June 2026) that it trains its own AI on de-identified audio recordings and transcriptions. Storage runs through cloud providers based in the United States, including Amazon Web Services.
- Fireflies.ai says it doesn't use personal information for model training, and processes data on servers "in the United States and other countries".
- Microsoft Teams stores recordings and transcripts in OneDrive and SharePoint and deletes them after 120 days by default. Teams falls under Microsoft's EU Data Boundary, with the exceptions Microsoft itself lists.
For an internal team meeting, that's often acceptable. For a conversation with a patient, a law firm's client, a job applicant or a customer calling about a dispute, it's a different matter, because those calls contain exactly the data you don't want sitting with a third party, often including health details, financial positions or the reasons someone is leaving a job. Why an EU region doesn't solve that is covered in EU data residency for AI.
Which speech models can transcribe Dutch locally?
In 2026 there are three serious open model families with Dutch support: Whisper from OpenAI, Voxtral from Mistral, and NVIDIA's Parakeet and Canary models. Accuracy is measured as word error rate (WER), the percentage of words that come out wrong. Lower is better. The figures below come from a single source, Mistral's Voxtral technical report (July 2025), so they're comparable with each other.
| Model | Licence | Size | Dutch WER, FLEURS | Dutch WER, Common Voice |
|---|---|---|---|---|
| Whisper large-v3 (OpenAI) | MIT | 1.55B parameters | 5.87% | 5.83% |
| Voxtral Mini (Mistral) | Apache 2.0 | 3B | 6.27% | 6.03% |
| Voxtral Small (Mistral) | Apache 2.0 | 24B | 4.94% | 3.98% |
| GPT-4o mini Transcribe (cloud, for reference) | closed | undisclosed | 5.54% | 6.47% |
Whisper large-v3 and OpenAI's cloud service are close on Dutch: on FLEURS the cloud is 0.3 points better, on Common Voice Whisper is 0.6 points better. Voxtral Small beats the cloud service on both benchmarks, but its model card says it needs around 55 GB of GPU memory. According to the Whisper GitHub page, large-v3 needs about 10 GB, and the faster turbo variant about 6 GB.
NVIDIA's Parakeet TDT 0.6B v3 and Canary-1B-v2 are two smaller models covering 25 European languages, Dutch included, under a CC-BY-4.0 licence. The WER figures on their model cards ignore capitalisation and punctuation, so they can't be compared one-to-one with the table.
For live transcription while the conversation is still running, Mistral released Voxtral Realtime in February 2026: an open 4-billion-parameter model under Apache 2.0 that runs on a single GPU with at least 16 GB of memory. Speed costs accuracy. On Dutch (FLEURS) it scores 7.07% at a 480-millisecond delay and 5.91% at 2.4 seconds. The more accurate batch version from the same family is only available through Mistral's API, not as an open model.
One caveat no benchmark shows: FLEURS and Common Voice are read-aloud sentences. A phone call with two people talking over each other, jargon and a bad line scores worse on every model. Test on ten of your own recordings before you choose.
How much compute does local transcription need?
Local transcription needs less compute than most people think. The faster-whisper benchmark, an optimised version of Whisper, transcribes 13 minutes of audio with the large-v2 model in 17 seconds on an RTX 3070 Ti with 8 GB of memory. That's a consumer card.
A worked example, indicative only. An advisory firm with 30 advisers who each hold two one-hour calls a day records 60 hours of audio a day. The benchmark gives 780 seconds of audio in 17 seconds, roughly 46 times faster than real time. Sixty hours of audio is 3,600 minutes. Divided by 46, that's about 78 minutes of compute a day on one card.
Speaker separation comes on top: who said what. The makers of the open pyannote diarization model report 31 seconds per hour of audio on an NVIDIA H100. They don't publish figures for smaller cards. At 60 hours of audio, that's 31 minutes on an H100, so even if a workstation card is five times slower it fits in an overnight run of just over two and a half hours. A local language model on the same machine writes the summary and action items. Which hardware fits is covered in AI hardware for business.
What may you record and transcribe under Dutch rules?
Recording a conversation you take part in isn't a criminal offence in the Netherlands. Articles 139a to 139c of the Dutch Criminal Code only criminalise recording by someone who isn't a participant and isn't acting on a participant's behalf. That's the criminal side. The GDPR applies on top.
The Dutch data protection authority, the Autoriteit Persoonsgegevens (AP), is specific about employees' phone calls on its page on monitoring communication tools. Recording is allowed where it's necessary. Telling employees once when they join is "not enough". The person on the other end of the line must be told in advance that the call is recorded and why. And recording covertly to coach or assess employees isn't allowed.
The works council's consent right
Transcribing client calls or internal meetings quickly falls under Article 27(1)(l) of the Dutch Works Councils Act: facilities "aimed at or suitable for" monitoring attendance, behaviour or performance. The word "suitable" carries weight. The AP points out that the works council also has a consent right over a system the employer doesn't use for monitoring but could. A searchable archive of every client call is such a system. How that process runs is covered in AI and works council consent.
Voice recognition and emotion recognition
Two features from commercial call analytics need extra care. The first is voice recognition. Speaker separation that tells "speaker 1" from "speaker 2" is not the same as a system that matches a voice to a known person using a stored voiceprint. That second kind processes biometric data as soon as it serves to uniquely identify someone (GDPR Article 4(14)), and Article 9's prohibition applies, with its narrow exceptions.
The second is emotion recognition. Article 5(1)(f) of the EU AI Act has banned AI systems that infer people's emotions in the workplace since 2 February 2025, except for medical or safety reasons. A call-centre dashboard that infers from an agent's voice whether they sound frustrated falls under it. With a local system, you decide which features are in it. With a SaaS package, you have to find them and switch them off.
When is a cloud note-taker fine?
For internal meetings without sensitive personal content, in an organisation already fully on Microsoft 365, Teams transcription is the shortest route. The transcript stays in your own tenant and expires after 120 days by default. Building your own transcription server doesn't pay off there. What an AI note-taker delivers for those meetings, and which tools exist, is in AI meeting notes.
Local transcription becomes the better choice for conversations under professional secrecy, calls involving health or financial data, job interviews, and client calls in volume. There, control over the recording outweighs the convenience of a subscription. The setup is straightforward: one server with a GPU, Whisper or Voxtral for the text, pyannote for the speakers, a local language model for the summary, and a link to your CRM or case system. How that fits a wider plan is in our guide to local AI for business.