Track 1: Unconstrained ASR Track
Goal: Advance the state of the art in dysarthric speech recognition.
Baseline Results
Official baselines follow the same ranking system as the HuggingFace Open ASR Leaderboard.
| Model ID | CER ↓ | WER ↓ | ||||||
|---|---|---|---|---|---|---|---|---|
| Dev | Test1 | Dev | Test1 | |||||
| All | Shared | Unshared | All | Shared | Unshared | |||
nvidia/canary-qwen-2.5b |
14.38 | 15.86 | 13.50 | 17.93 | 20.25 | 22.60 | 18.89 | 24.54 |
nvidia/parakeet-tdt-0.6b-v2 |
13.43 | 16.47 | 11.61 | 16.07 | 19.59 | 24.04 | 17.00 | 22.83 |
openai/whisper-large-v2 |
13.27 | 16.40 | 11.39 | 15.19 | 18.70 | 22.87 | 16.27 | 21.46 |
xiuwenz2/whisper-large-v2-ft-SAP-0430-pd-merged-ct2 |
9.10 | 8.79 | 9.29 | 13.49 | 13.04 | 12.12 | 13.58 | 19.05 |
Note: The baseline results shown here may differ slightly from Codabench Leaderboard due to evaluation-environment differences.