ark-asr-3b-open-asr-leaderboard-results

# ARK-ASR-3B Open ASR Leaderboard Results Raw JSONL manifests for `AutoArk-AI/ARK-ASR-3B` on the public English short-form `hf-audio/open-asr-leaderboard` splits. These manifests were generated on a local 8x RTX 4090 machine and scored with the shared Open ASR Leaderboard scorer: ```bash PYTHONPATH=. python - <<'PY' from normalizer.eval_utils import score_results score_results( 'ark_asr/results.AutoArk-AI-ARK-ASR-3B_20260622_official', 'AutoArk-AI/ARK-ASR-3B', ) PY ``` Important: `earnings22` uses the official `hf-audio/open-asr-leaderboard` leaderboard data, not the old chunked `distil-whisper/earnings22` cache. ## Results | Split | Samples | WER | RTFx | | --- | ---: | ---: | ---: | | ami/test | 11653 | 8.91 | 272.25 | | earnings22/test | 2737 | 8.25 | 243.20 | | gigaspeech/test | 19898 | 7.30 | 140.23 | | librispeech/test.clean | 2620 | 1.09 | 293.56 | | librispeech/test.other | 2939 | 2.41 | 284.61 | | spgispeech/test | 39341 | 2.49 | 208.87 | | voxpopuli/test | 1842 | 5.48 | 293.41 | Composite: - Average WER: `5.13` - Overall RTFx: `197.14` ## Files ```text MODEL_AutoArk-AI-ARK-ASR-3B_DATASET_hf-audio-open-asr-leaderboard_ami_test.jsonl MODEL_AutoArk-AI-ARK-ASR-3B_DATASET_hf-audio-open-asr-leaderboard_earnings22_test.jsonl MODEL_AutoArk-AI-ARK-ASR-3B_DATASET_hf-audio-open-asr-leaderboard_gigaspeech_test.jsonl MODEL_AutoArk-AI-ARK-ASR-3B_DATASET_hf-audio-open-asr-leaderboard_librispeech_test.clean.jsonl MODEL_AutoArk-AI-ARK-ASR-3B_DATASET_hf-audio-open-asr-leaderboard_librispeech_test.other.jsonl MODEL_AutoArk-AI-ARK-ASR-3B_DATASET_hf-audio-open-asr-leaderboard_spgispeech_test.jsonl MODEL_AutoArk-AI-ARK-ASR-3B_DATASET_hf-audio-open-asr-leaderboard_voxpopuli_test.jsonl ```