// Announcement · Medical speech

The lowest published Dutch speech-recognition error rate — on-premise, and independently checkable.

On the public Dutch benchmarks, the speech engine behind Speakeasy posts the lowest word error rate we are aware of — and every number is scored with the benchmark's own tooling and released for anyone to re-run.

CodeSpark Tech · Amsterdam

#1 on the Dutch community benchmark

Scored with the maintainers' official NIST-sclite tooling, our engine reaches 3.5% WER on the Common Voice 17 Dutch test set, against 4.3% for the previous best entry on the community board. A same-model control run reproduces that 4.3% exactly — so the improvement is a real accuracy gain, not a decoding trick.

Ahead of the newest multilingual models

On the HuggingFace Open ASR harness, the same engine outperforms NVIDIA's Canary-1B-v2 and Parakeet-TDT-0.6b-v3 on all three Dutch test sets — FLEURS, Common Voice and MLS.

Benchmark (official scorer)Our engineBest alternativeStanding
Common Voice NL — NIST-sclite3.5%4.3%#1 on the Dutch community board
FLEURS NL — HF Open ASR4.635.27ahead of NVIDIA Canary-1B-v2
Common Voice NL — HF Open ASR4.305.47ahead of NVIDIA Parakeet-0.6b-v3
MLS NL — HF Open ASR5.119.64ahead of both
Medicijnjournaal — held-out medical11.4%12.8%beats trained baselines, no medical data

Word error rate, lower is better. Each figure is scored with the venue's own official tooling.

Strong on medical speech, with no medical training data

On the held-out Medicijnjournaal corpus from the Radboud HoMed project, it reaches 11.4% WER — below the trained Wav2vec2.0 (12.8%) and Kaldi-NL (16.1%) baselines — despite never seeing medical data in training.

Built to run where the cloud can't

The engine runs fully on-premise or air-gapped, so audio never leaves the customer's building, with hybrid post-quantum encryption for data at rest. That is a hard requirement in Dutch healthcare, mental-health care and government that hosted engines structurally cannot meet.

Receipts, not marketing

Every figure is scored with the venue's own tooling and published with per-utterance output, a same-model control, and a train-on-test contamination check — so anyone can re-derive them in an afternoon.

Independently verifiable. The per-utterance benchmark output, contamination check and control run are public on GitHub and Hugging Face. CodeSpark Tech is a member of the NVIDIA Inception program for AI startups.

The engine is fine-tuned from an openly-licensed base model (Apache-2.0) on rights-clean public Dutch corpora; the differentiation is the training and decode-time pipeline and the on-premise deployment stack. Benchmark comparisons cite the leading published open-weight and hosted systems on each set.