Everyone scores well in the studio.
The consulting room is the real test. That is where the rest fall over — publicly measured, not claimed by us.
Dutch speech · your hardware · your key
Nothing leaves
Nothing leaves
the building.
The sharpest Dutch speech model publicly measured to date — running on your own rack, behind your own key.
Twenty minutes on your own audio. Answer within one working day.
● Recording · local
16 kHz · nl-NL
Seal of record—
Hash9f2c·b710·4e83·a41d
SignedML-DSA-65 · FIPS 204
Keylocal · HSM vault
Outbound traffic0 bytes
Procurement-ready
NEN 7510 · designed to the standard
BIO2 · Dutch public-sector framework
GDPR · DPIA support
EU data residency
Processing agreement ready
Open foundation · export without lock-in
Dutch-owned · KVK 42053556
01 — Where it counts
Wherever every word counts.
The Hague
The room where decisions are made
Consultations, hearings, decision-making. Every record traceable to what was said — speaker recognition, time codes and signed hashes, inside your own walls.
The consultation
Attention for the person, not the screen
From conversation to record while the contact continues. Structured output, ready for the system that is already there.
The study
Recording without getting in between
Record now, finish later. No live text in the conversation, but a complete and verifiable trail afterwards.
What is said in the Netherlands, stays in the Netherlands.
02 — Three words, three numbers
3,5%
Provable
WER on Common Voice 17.0 NL. Official NIST scoring, full output public — checkable without us.
0
Sovereign
Bytes leaving your walls with on-premise deployment. No outbound connection to trust.
PQC
Quantum-resistant
ML-KEM-1024 and ML-DSA-65, the final NIST standards. Intercepted today is still sealed in ten years.
03 — The wall
A supplier can be acquired. Your installation cannot.
Dutch ownership is a snapshot; architecture is permanent. VeraDossier runs entirely inside your own walls — no call-home, no licence server, no dependency that changes hands tomorrow.
Inside your walls
- Audio, transcript and record
- Model, weights and inference
- Keys in your own HSM vault
- Logging, audit and retention
What goes out
0
No telemetry, no update server, no crash report. There is no outbound connection to trust.
Studio audio · Common Voice 17.0 NL
Same test set, same scoring tool as the public board. All 11,266 clips. Lower is better.
Measured August 2026 against the versions published at that time. Models named are trademarks of their owners; the comparison uses the same audio and the same scoring tool and can be recomputed via Method.
The consulting room · real clinical audio
Generic models lose heavily on clinical audio. Of the systems with a published score on this benchmark, ours is the lowest — full sclite output public.
The proof nobody gives
Read our
mistakes.
Every vendor publishes a percentage. We publish the sentences. Same audio, same scoring tool, three models side by side — and the clips where we do worse are right there with them.
69.8%
of sentences exactly right — no correction needed at all. Canary 64.7% · Parakeet 65.6%.
19 : 2
clips where an NVIDIA model collapses and we do not, against the reverse. The outliers do the damage, not the average.
5
clips completely wrong (100% WER or more). Canary 9 · Parakeet 11. In a record, that is the line that hurts.
← → to browse · 1-4 to select
Where the comparison is uneven
Our model is fine-tuned on the Common Voice training split, while the board figure of 4.3% was achieved zero-shot. That is an advantage on this particular test set. We checked for leakage — 9 of the 11,266 test clips (0.08%), worth 0.1 percentage points. The medical benchmark is fully unseen.
Where we lose
Speed. On the same harness Parakeet reaches sixteen times our throughput — 1557 against 94 RTFx. Partly the test setup, mostly real: 1.55 billion parameters against 0.6 billion.
Check it yourself
- 00Methode en verantwoording → The full method: which test sets, which scoring tool, which checks — and what we do not claim.
- 01ASR_NL_results · PR #7 ↗ The submission itself: Common Voice 3.5% and Medicijnjournaal 11.4%, with sclite output, the control run and the leakage analysis. Open with the volunteers who maintain the board.
- 02HuggingFace · modelkaart ↗ Evaluation results, the model lineage and the same caveats as above. Weights are commercial, the evaluation is open.
- 03Open ASR Leaderboard · issue #187 ↗ Our proposal for a Dutch track, including the head-to-head against NVIDIA and the speed concession.
An answer for everyone at the table.
Nobody decides on speech recognition alone. The board, the DPO, the CISO, procurement and application management each ask their own question. Here are the answers — short, and verifiable.
Who can get at our data?
NL·EU
Only the Netherlands and the EU. No foreign law can demand what was never sent — on-premise, nothing leaves your building.
Do you train on our recordings?
Never
Off by default, only with written opt-in, contractually fixed. The data processing agreement is ready; DPIA support is included.
Who holds the key?
You
Client-side encrypted, keys in your own HSM vault, rotated every 90 days. Post-quantum: ML-KEM-1024 and ML-DSA-65, the final NIST standards.
And if we want to stop?
4open formats
Records in DOCX, TXT, SRT and VTT. Open foundation (Whisper MIT · Dutch fine-tune Apache-2.0). Stopping never costs you your archive.
Does it fit our EHR?
76%NL · ChipSoft HiX
Structured SOAP and GIZ output today, ready for HiX, Epic and Nedap ONS. HiX integration in development, Kidos · KD+ · BDS in pilot — status on request.
06 — Procurement
Procurement questions.
Procurement questions.
No hedging.
What every buyer asks before anything gets signed — including the answers where we are not there yet.
01Do you train on our audio?
No. Your audio and transcripts are never used for training without explicit written opt-in.
02Where is our data stored?
On-premise, solely on your own hardware: nothing goes out. Choose the hosted variant and processing sits in EU data centres with key management still yours. Only on-premise guarantees we never see unencrypted audio.
03Does this integrate with our EHR/HIS?
Structured SOEP/GIZ output. Integrations with youth-care EHRs — Kidos, KD+ and BDS v4 — are in pilot with selected practices; ask about the status for your system.
04What is the status of certifications?
We hold no certificate today. The environment is designed and operated to NEN 7510 and BIO2; formal certification and the independent audits are in preparation. We would rather write that down ourselves than have you find it in the audit. Documentation on request.
05What does it cost?
Custom pricing per practice/organization — a proposal within one working day after a short demo.
06Can I verify the benchmark figures myself?
Yes. Every figure is scored with the official ASR_NL benchmark tooling (NIST sclite) and the full output is public in the community results repository — including our train-on-test check and a control run of the best open model on exactly the same clips.
07We build care software ourselves — can the speech engine be licensed?
Yes. The Vlotspraak engine is licensable per tenant and runs entirely inside your infrastructure or your customer's. We validate on your own test data under NDA first — if the engine does not beat your current solution, it cost you one hour.
Does it hold up on our audio?
20minutes
A demo on your own recordings, a proposal within one working day. If it disappoints, it cost you twenty minutes.
Systems and models named here are trademarks of their respective owners; naming them describes interoperability and implies no partnership or endorsement. VeraDossier is documentation software, not a medical device: its output supports record-keeping and does not replace professional or clinical judgement. Human final review remains required.