jyoti / evaluation.json
voidash's picture
D3 (step 100000) becomes the main codec.pt; C moved to variants/C-noise-p050; card: D3 gates, conversational Nepali (same-domain caveat), open D3-vs-D3C question, D3 training data and attributions
2d7d9d7 verified
Raw History Blame Contribute Delete
4.03 kB
{
"schema": "jyoti-public-evaluation-v1",
"arm": "D3-manifest50-extra-p050",
"inference_checkpoint_sha256": "417ed68677b85e644dfeea2415f4d0349322861f7201dcbd9d16435f18a04e00",
"training_checkpoint_sha256": "c4fe5a6cfe741da64ff74c928a05189c8a62ab6f3556d72042b404eb9e136b9d",
"step": 100000,
"quality_qualified": false,
"sealed_final_cohort_opened": false,
"english": {
"asr": "facebook/wav2vec2-base-960h @ 22aad52d435eb6dbaf354bdad9b0da84ce7d6156 (greedy CTC)",
"pools": {
"clean": "128 LibriSpeech dev-clean utterances",
"difficult": "128 LibriSpeech dev-other utterances",
"pink_10db": "the clean pool + synthetic pink noise at 10 dB SNR (codec input)",
"pink_5db": "the clean pool + synthetic pink noise at 5 dB SNR"
},
"mode": "packed bitstream (encode -> bytes -> decode)",
"results": {
"clean": {
"codec_wer": 0.06684611933602512,
"original_wer": 0.02826379542395693,
"codec_cer": 0.02497897392767031,
"original_cer": 0.007485281749369218,
"wer_increase_pp": 3.8582323912068186,
"mean_stoi": 0.8995658662229111,
"mean_pesq_wb": 2.111289313994348,
"utterances": 128
},
"difficult": {
"codec_wer": 0.20997732426303856,
"original_wer": 0.07936507936507936,
"codec_cer": 0.09342347879532882,
"original_cer": 0.029765563262797435,
"wer_increase_pp": 13.061224489795919,
"mean_stoi": 0.8848347751522835,
"mean_pesq_wb": 1.973732240498066,
"utterances": 128
},
"pink_10db": {
"codec_wer": 0.24136384028712426,
"original_wer": 0.16913414087034545,
"codec_cer": 0.11648444070647603,
"original_cer": 0.07846930193439866,
"wer_increase_pp": 7.222969941677881,
"mean_stoi": 0.7282501492237068,
"mean_pesq_wb": 1.2500946335494518,
"utterances": 128
},
"pink_5db": {
"codec_wer": 0.4064602960969044,
"original_wer": 0.5118887393449978,
"codec_cer": 0.21412952060555088,
"original_cer": 0.30538267451640033,
"wer_increase_pp": -10.542844324809336,
"mean_stoi": 0.6389370863985571,
"mean_pesq_wb": 1.1396433534100652,
"utterances": 128
}
}
},
"development_gates": {
"note": "Engineering targets registered before training; not values from any paper. Unchanged.",
"gates": [
{
"gate": "clean_wer_increase_pp",
"value": 3.8582323912068186,
"operator": "<=",
"threshold": 5.0,
"passed": true
},
{
"gate": "difficult_wer_increase_pp",
"value": 13.061224489795919,
"operator": "<=",
"threshold": 10.0,
"passed": false
},
{
"gate": "pink_10db_wer_increase_pp",
"value": 7.222969941677881,
"operator": "<=",
"threshold": 10.0,
"passed": true
},
{
"gate": "pink_5db_wer_increase_pp",
"value": -10.542844324809336,
"operator": "<=",
"threshold": 10.0,
"passed": true
},
{
"gate": "clean_mean_stoi",
"value": 0.8995658662229111,
"operator": ">=",
"threshold": 0.85,
"passed": true
}
],
"passed": 4,
"total": 5
},
"nepali_supplemental": {
"asr": "ampixa/nepali-conformer-offline @ a7830b7ab16fbc2659b590b4e5e4030737c4caf9",
"pool": "128 SLR54 utterances from 32 held-out development speakers (not in training)",
"codec": {
"word_errors": 242,
"reference_words": 378,
"character_errors": 596,
"reference_characters": 2106,
"wer": 0.6402116402116402,
"cer": 0.28300094966761635
},
"original_audio": {
"word_errors": 154,
"reference_words": 378,
"character_errors": 310,
"reference_characters": 2106,
"wer": 0.4074074074074074,
"cer": 0.14719848053181386
}
},
"sources": {
"english_summary_sha256": "15f08d1ee80caa41b4a95ae5a473d5196205263f99e1ec089b6256eff8fe7abe",
"nepali_summary_sha256": "996f741f1c5c816b618503a00bc3140bbb246cbe22ef1f5e73352cab00496b1c",
"nepali_original_summary_sha256": "80f7fef5e23d2e5932d0de600c8268407d81727fda3f6592923b6a664de49d38",
"qualification_targets_sha256": "23e14e40b027de87c217488ee2c7989176e2930277ecc991395de5ef845b5145"
}
}