oruk-loop-153M / results /evaluation.json
NathanRoll's picture
oruk-loop-153M: weights, inference code, model card, licenses, results, media
58b0065 verified
Raw History Blame Contribute Delete
10.1 kB
{
"model": "oruk/oruk-loop-153M",
"stored_parameters": 153202638,
"parameters_by_part": {
"audio_encoder_loop": 41213952,
"decoder_loop": 67886080,
"token_table": 28932430,
"kept_qwen3_asr_modules": 15170176
},
"teacher": {
"model": "Qwen/Qwen3-ASR-1.7B-hf",
"stored_parameters": 2038052480
},
"scoring": "Whisper EnglishTextNormalizer on reference and hypothesis; empty / \"ignore time segment in scoring\" references dropped; corpus WER per set; audio >30 s split at the quietest 100 ms frame in the last 5 s before each 30 s limit",
"test_sets": {
"leaderboard": "hf-audio/open-asr-leaderboard (original transcripts)",
"tedlium": "sanchit-gandhi/tedlium-data test"
},
"evaluation_pipeline": {
"beam4": {
"decoding": "beam search, 4 beams, length_penalty 1.0, bf16",
"sets": {
"ami": {
"wer_pct": 11.222,
"reference_words": 89564,
"utterances": 11653,
"dropped": 990,
"hours": 8.68
},
"earnings22": {
"wer_pct": 9.804,
"reference_words": 48899,
"utterances": 2737,
"dropped": 4,
"hours": 5.43
},
"gigaspeech": {
"wer_pct": 10.455,
"reference_words": 403656,
"utterances": 19898,
"dropped": 33,
"hours": 35.36
},
"ls_clean": {
"wer_pct": 2.399,
"reference_words": 53027,
"utterances": 2620,
"dropped": 0,
"hours": 5.4
},
"ls_other": {
"wer_pct": 5.953,
"reference_words": 52882,
"utterances": 2939,
"dropped": 0,
"hours": 5.34
},
"spgispeech": {
"wer_pct": 2.476,
"reference_words": 969328,
"utterances": 39341,
"dropped": 0,
"hours": 100.0
},
"tedlium": {
"wer_pct": 3.772,
"reference_words": 28207,
"utterances": 1155,
"dropped": 314,
"hours": 3.06
},
"voxpopuli": {
"wer_pct": 7.546,
"reference_words": 44079,
"utterances": 1842,
"dropped": 0,
"hours": 4.93
}
},
"mean_7_leaderboard_sets": 7.122,
"mean_8_incl_tedlium": 6.703
},
"greedy": {
"decoding": "greedy, bf16",
"sets": {
"ami": {
"wer_pct": 11.575,
"reference_words": 89564,
"utterances": 11653,
"dropped": 990,
"hours": 8.68
},
"earnings22": {
"wer_pct": 10.289,
"reference_words": 48899,
"utterances": 2737,
"dropped": 4,
"hours": 5.43
},
"earnings22_cleaned_aa": {
"wer_pct": 10.916,
"reference_words": 15418,
"utterances": 6,
"dropped": 0,
"hours": 1.92
},
"gigaspeech": {
"wer_pct": 10.75,
"reference_words": 403656,
"utterances": 19898,
"dropped": 33,
"hours": 35.36
},
"ls_clean": {
"wer_pct": 2.529,
"reference_words": 53027,
"utterances": 2620,
"dropped": 0,
"hours": 5.4
},
"ls_other": {
"wer_pct": 6.322,
"reference_words": 52882,
"utterances": 2939,
"dropped": 0,
"hours": 5.34
},
"spgispeech": {
"wer_pct": 2.607,
"reference_words": 969328,
"utterances": 39341,
"dropped": 0,
"hours": 100.0
},
"tedlium": {
"wer_pct": 3.793,
"reference_words": 28207,
"utterances": 1155,
"dropped": 314,
"hours": 3.06
},
"voxpopuli": {
"wer_pct": 7.698,
"reference_words": 44079,
"utterances": 1842,
"dropped": 0,
"hours": 4.93
},
"voxpopuli_cleaned_aa": {
"wer_pct": 4.611,
"reference_words": 17784,
"utterances": 628,
"dropped": 0,
"hours": 1.98
}
},
"mean_7_leaderboard_sets": 7.396,
"mean_8_incl_tedlium": 6.945
},
"note_cleaned_sets": "earnings22_cleaned_aa / voxpopuli_cleaned_aa: ArtificialAnalysis re-transcribed versions of the same test audio (ArtificialAnalysis/Earnings22-Cleaned-AA, hf-audio voxpopuli_cleaned_aa); not part of the leaderboard average"
},
"out_of_domain_beam4": {
"note": "fleursu = google/fleurs en_us test (647); cvu = Common Voice test from hf-audio/open-asr-leaderboard, fixed random 4000 (seed 20261014); edacc = edinburghcstr/edacc test; peoples = MLCommons/peoples_speech clean test, fixed random 4000 (3,999 scored)",
"sets": {
"ls_clean": {
"wer_pct": 2.399,
"reference_words": 53027,
"utterances": 2620,
"dropped": 0,
"hours": 5.4
},
"fleursu": {
"wer_pct": 9.882,
"reference_words": 14279,
"utterances": 647,
"dropped": 0,
"hours": 1.77
},
"cvu": {
"wer_pct": 17.148,
"reference_words": 38629,
"utterances": 4000,
"dropped": 0,
"hours": 6.62
},
"edacc": {
"wer_pct": 20.158,
"reference_words": 157116,
"utterances": 8537,
"dropped": 752,
"hours": 16.94
},
"peoples": {
"wer_pct": 20.154,
"reference_words": 70861,
"utterances": 3999,
"dropped": 1,
"hours": 6.93
}
}
},
"frontier_analysis": {
"snapshot": "hf-audio/open_asr_leaderboard English results, fetched 2026-10-14",
"previous_frontier_points": [
[
"nvidia/parakeet-tdt_ctc-110m",
0.11,
7.977
],
[
"nvidia/canary-180m-flash",
0.18,
7.484
],
[
"usefulsensors/moonshine-streaming-medium",
0.24,
7.181
],
[
"nvidia/parakeet-tdt-0.6b-v2",
0.6,
6.436
]
],
"previous_frontier_at_153M_pct": 7.645,
"gap_at_153M_pct_points": 0.524,
"size_where_previous_frontier_reaches_7.12_B": 0.2587,
"parameter_ratio_at_equal_accuracy": 1.69,
"method": "linear interpolation of 7-set mean WER in log10(size) between neighbouring frontier models",
"note": "gap uses the exact beam-4 mean 7.122"
},
"release_package_rescore": {
"code": "this repository: AutoModel.from_pretrained(trust_remote_code=True) + model.transcribe, batch 64 (beam: 32), one A100 40GB per job",
"greedy": {
"sets": {
"ami": {
"wer_pct": 11.548,
"reference_words": 89564,
"utterances": 11653,
"dropped": 990,
"hours": 8.68,
"seconds": 308
},
"earnings22": {
"wer_pct": 10.27,
"reference_words": 48899,
"utterances": 2737,
"dropped": 4,
"hours": 5.43,
"seconds": 162
},
"tedlium": {
"wer_pct": 3.907,
"reference_words": 28207,
"utterances": 1155,
"dropped": 314,
"hours": 3.06,
"seconds": 120
},
"ls_clean": {
"wer_pct": 2.542,
"reference_words": 53027,
"utterances": 2620,
"dropped": 0,
"hours": 5.4,
"seconds": 129
},
"ls_other": {
"wer_pct": 6.431,
"reference_words": 52882,
"utterances": 2939,
"dropped": 0,
"hours": 5.34,
"seconds": 152
},
"voxpopuli": {
"wer_pct": 7.709,
"reference_words": 44079,
"utterances": 1842,
"dropped": 0,
"hours": 4.93,
"seconds": 103
},
"gigaspeech": {
"wer_pct": 10.762,
"reference_words": 403656,
"utterances": 19898,
"dropped": 33,
"hours": 35.36,
"seconds": 1021
},
"spgispeech": {
"wer_pct": 2.607,
"reference_words": 969328,
"utterances": 39341,
"dropped": 0,
"hours": 100.0,
"seconds": 2221
}
},
"mean_7_leaderboard_sets": 7.41,
"mean_8_incl_tedlium": 6.972
},
"beam4": {
"sets": {
"ami": {
"wer_pct": 11.224,
"reference_words": 89564,
"utterances": 11653,
"dropped": 990,
"hours": 8.68,
"seconds": 635
},
"earnings22": {
"wer_pct": 9.802,
"reference_words": 48899,
"utterances": 2737,
"dropped": 4,
"hours": 5.43,
"seconds": 374
},
"tedlium": {
"wer_pct": 3.772,
"reference_words": 28207,
"utterances": 1155,
"dropped": 314,
"hours": 3.06,
"seconds": 238
},
"ls_clean": {
"wer_pct": 2.404,
"reference_words": 53027,
"utterances": 2620,
"dropped": 0,
"hours": 5.4,
"seconds": 267
},
"ls_other": {
"wer_pct": 5.942,
"reference_words": 52882,
"utterances": 2939,
"dropped": 0,
"hours": 5.34,
"seconds": 293
},
"voxpopuli": {
"wer_pct": 7.543,
"reference_words": 44079,
"utterances": 1842,
"dropped": 0,
"hours": 4.93,
"seconds": 222
},
"gigaspeech": {
"wer_pct": 10.446,
"reference_words": 403656,
"utterances": 19898,
"dropped": 33,
"hours": 35.36,
"seconds": 2156
},
"spgispeech": {
"wer_pct": 2.475,
"reference_words": 969328,
"utterances": 39341,
"dropped": 0,
"hours": 100.0,
"seconds": 4312
}
},
"mean_7_leaderboard_sets": 7.119,
"mean_8_incl_tedlium": 6.701
}
},
"teacher_on_our_harness": {
"model": "Qwen/Qwen3-ASR-1.7B-hf",
"decoding": "greedy, bf16",
"sets": {
"spgispeech": {
"wer_pct": 2.835,
"reference_words": 969328,
"utterances": 39341,
"dropped": 0,
"hours": 100.0,
"seconds": 1392
},
"tedlium": {
"wer_pct": 2.28,
"reference_words": 28207,
"utterances": 1155,
"dropped": 314,
"hours": 3.06,
"seconds": 58
},
"ls_clean": {
"wer_pct": 1.618,
"reference_words": 53027,
"utterances": 2620,
"dropped": 0,
"hours": 5.4,
"seconds": 85
},
"ami": {
"wer_pct": 10.481,
"reference_words": 89564,
"utterances": 11653,
"dropped": 990,
"hours": 8.68,
"seconds": 420
},
"earnings22": {
"wer_pct": 10.062,
"reference_words": 48899,
"utterances": 2737,
"dropped": 4,
"hours": 5.43,
"seconds": 101
},
"gigaspeech": {
"wer_pct": 8.738,
"reference_words": 403656,
"utterances": 19898,
"dropped": 33,
"hours": 35.36,
"seconds": 632
},
"ls_other": {
"wer_pct": 3.364,
"reference_words": 52882,
"utterances": 2939,
"dropped": 0,
"hours": 5.34,
"seconds": 92
},
"voxpopuli": {
"wer_pct": 6.237,
"reference_words": 44079,
"utterances": 1842,
"dropped": 0,
"hours": 4.93,
"seconds": 68
}
},
"mean_7_leaderboard_sets": 6.191,
"mean_8_incl_tedlium": 5.702,
"leaderboard_snapshot_7set_mean": 6.25
}
}