Spaces:

---

tvosch

/

Dutch-ASR-Leaderboard

like 7

Running

Add CodeSparkTech Vlotspraak Lange Jaap v0.7 (proprietary Dutch ASR)

Results for CodeSparkTech Vlotspraak Lange Jaap v0.7, an on-premise commercial Dutch ASR model by CodeSpark Technologies B.V. (Whisper large-v3 architecture, 1.55B params), produced with the leaderboard's own asr-nl-eval harness (default normalizer, zero failed samples).

Full per-sample reference/hypothesis pairs are included in the results JSON for verification (same basis as the existing ElevenLabs / Resonate / Murmel proprietary entries).

Notes

• No training on any test split — standard fine-tune on public Dutch training data; test splits untouched.
• No decode-time language model — pure acoustic model decoded through your harness.
• Weights are proprietary and not publicly released; registered under the proprietary toggle in app.py.
• Benchmarked on an RTX 5090 (you benchmark local models on H100), so RTF figures are indicative; WER is hardware-independent.
• The CommonVoice 25 NL column depends on the config in tvosch/Dutch-ASR-leaderboard#1 (your roadmap item); the result rides in the JSON and renders once that column exists.

Happy to adjust the display name or formatting to your conventions.