Cadence™ 1 · In development
Cadence™
Speech recognition for the languages of Europe. Built for real conversations, phone calls and meetings, and small enough to run on your own computer.
00:00:04.2
Goedemorgen, ik bel over de afspraak van donderdag. Kunnen we die verzetten naar volgende week?
Illustration. Cadence is in development.
The best open speech models get about five words in every hundred wrong when someone reads Dutch aloud. In a real conversation, on the phone, with an accent or a child speaking, they do much worse. That's the speech we want to get right.
Today's best on Dutch
Word error rate on the FLEURS Dutch test set: read sentences, scored on how many words a model gets wrong. These are published results from other models, and the bar Cadence has to clear.
Source: published FLEURS results, November 2026. Read speech is the easy case; studies of Dutch conversational and phone speech report much higher error rates for the same models.
| Model | Size | Error rate |
|---|---|---|
| Voxtral Small Mistral | 24 B | 5.1% |
| Whisper large-v3 OpenAI | 1.5 B | 5.4% |
| Whisper large-v3 turbo OpenAI | 0.8 B | 6.0% |
| Canary 1B v2 NVIDIA | 1 B | 6.3% |
| Cadence™ 1 Ordnary · target | < 1 B | 5.5% or better |
What we're building
- Dutch first
- Trained and measured on Dutch before anything else, including Flemish and non-native speakers. More European languages follow.
- < 1 B
- Under a billion parameters: aimed at the accuracy of models up to twenty times its size, at a fraction of the compute.
- CPU
- Small enough to transcribe on an ordinary computer, so recordings of meetings, calls and patients never have to leave it.
- Open
- Released with open weights and a model card, evaluated on public benchmarks anyone can rerun.
Roadmap
No release date. Each step happens when the one before it is done.
- 01 Done
Baseline
Published results for today's best open models on Dutch, collected and compared. This is the bar to clear.
- 02 Next
Train on Dutch
Continued training of an open, commercially licensed speech model on public Dutch speech: Common Voice, VoxPopuli and Multilingual LibriSpeech.
- 03 Planned
Measure and release
Released only if it matches the best open models on Dutch, with a model card, open weights and every score published.
- 04 Later
Conversations and more languages
Spontaneous speech, phone calls and meetings, where today's models struggle most, and the other European languages.
Questions
How we build AIWhen is it available?
When it's good enough. Cadence is released once it matches the best open models on Dutch in public benchmarks, not on a date.
Why build another speech model?
The models that do well on Dutch are either very large and expensive to run, or noticeably worse on conversations, accents and phone audio. We want one that's small enough to run anywhere and good at the speech people actually produce.
Will my recordings be used for training?
No. Cadence is trained on public datasets with licenses that allow it. Recordings you transcribe with it stay on your own systems.