File size: 5,992 Bytes
63b692e ab5ae4a 63b692e ab5ae4a 63b692e ab5ae4a 63b692e ab5ae4a 63b692e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 | {
"model_type": "aria_linear_head",
"model_version": "v0.2.3.3",
"status": "formal cold-start release; G3 PASS under revision R1; T_diag uncalibrated (pending ruling); §2.5 locked benchmark run 2026-10-18, not significant vs. the v1 frozen CAPE-V severity head",
"contract": {
"input": "segment A: 16 kHz mono waveform, or a raw recording (stage 0 is bundled)",
"output": "abi_prime [0,10], y_hat_mm = 10 * abi_prime (PVQD CAPE-V overall severity, mm), s [0,100]",
"validated_input": "PVQD protocol compound recordings (sustained /a/, /i/ + six CAPE-V sentences), whole file through stage 0",
"out_of_scope": "any-recording vowel cutting; the front end that locates and cuts a sustained vowel does not exist yet",
"health_to_severity_mapping": "none; abi_prime is CAPE-V severity / 10 directly"
},
"pipeline": {
"chain": "stage 0 segment A -> native 9 features -> L3 linear head -> clip [0,10] (G6 identity) -> g table -> S",
"seg_ver": "stage0-v0.2.1.2",
"feature_engine": "native",
"head": "single affine linear regression (refit L3, lambda = 100)",
"feature_names": [
"cpps",
"jitter_local",
"gne_max",
"hfno6000",
"hnrd",
"h1h2",
"shimmer_local_db",
"shimmer_local",
"psd"
],
"abi_prime_clip": [
0.0,
10.0
],
"g_ver": 3,
"g_space": "ABI_prime_corr [0,10] -> S [0,100]",
"g6_correction_active": false
},
"package_layout": {
"weights/model.json": "L3 coefficients, clip, g version, cold-start thresholds, dev metrics, package identity",
"weights/g_table.json": "dimension chain g constant table (knots / p / sigma) with its acceptance record",
"inference/predict.py": "entry point: --audio | --segment-a | --requests",
"inference/stage0/": "stage 0, seg_ver stage0-v0.2.1.2 (13 modules, byte-identical to the frozen prediction run)",
"inference/native_features/": "native nine-feature engine (5 modules, byte-identical to the frozen prediction run)",
"history/v0.1-pre/": "archived v0.1-pre package (card, weights, entry point)"
},
"dependencies": {
"python": ">=3.10",
"core": [
"numpy",
"scipy",
"soundfile",
"onnxruntime"
],
"segment_a_only_path_needs": [
"numpy",
"scipy"
],
"vad_model": {
"name": "SileroVAD.onnx (Silero v5)",
"sha256": "1a153a22f4509e292a94e67d6f9b85e8deb25b4988682b7e174c65279d8788e3",
"bundled": false,
"resolution": "--vad-model PATH, $ARIA_VAD_MODEL, or inference/models/SileroVAD.onnx"
}
},
"training_data": {
"corpus": "PVQD",
"fit_pool_rows": 212,
"holdout_rows": 12,
"category_distribution_fit_pool": {
"normal": 84,
"mild": 79,
"moderate": 31,
"severe": 18
},
"label": "PVQD CAPE-V overall severity average / 10",
"caliber": "relaxed / training (R1 and W_窗数 not excluded)",
"locked_60_used_in_fit": false,
"test_labels_in_package": false,
"source_repo": "aria-training"
},
"evaluation": {
"dev": {
"holdout_n": 12,
"holdout_rs": 0.8881118881118882,
"holdout_rs_L0": 0.8951048951048951,
"holdout_mae_abi": 1.254975337236816,
"cv_mean_rs": 0.64148287217087,
"cv_rs_range": 0.20672451853288648,
"g3_refit_gate": "PASS under revision R1 (tau = 0.369)",
"g3_refit_gate_original": "FAIL (CV range > 0.10; not-worse-than-L0 -0.007)"
},
"locked_benchmark": {
"run": "§2.5, 2026-10-18, one-shot",
"comparator": "v1 frozen CAPE-V severity head (native mm, not recomputed)",
"paired_n": 57,
"primary_endpoint_median_d_mm": 0.693,
"primary_endpoint_ci95_mm": [
-2.408,
3.012
],
"wilcoxon_one_sided_p": 0.790005,
"sign_test_one_sided_p": 0.395683,
"verdict": "not significant; do not claim v0.2.3.3 beats the v1 severity head on the locked set",
"aria_rs": 0.7164,
"aria_rs_ci95": [
0.5338,
0.8429
],
"aria_mae_mm": 11.971,
"aria_mae_ci95_mm": [
9.553,
14.482
],
"v1_rs": 0.813,
"v1_mae_mm": 10.951,
"strata_descriptive": "Normal/Mild comparable; Moderate/Severe descriptively weaker for v0.2.3.3",
"disclosures": {
"intent_to_screen": "60 allocated -> 58 scored -> 2 VAD rejections (PT084, PT134)",
"rejected_in_package_caliber": "PT097 rejected:R2|R5 (no prediction produced)",
"product_caliber": "22 scored / 36 rejected:R1, descriptive only, not part of any endpoint"
},
"frozen_predictions_sha256": "3ff22e6a55b13673b17d7d02ba5ac564ca777de96e927186172991508b144d69",
"generation_manifest_sha256": "08655d7bdf733f2d6ea138008f7ce10076e55017de60582058a2da2fe6002c38",
"report": "aria-training/docs/v0233-locked-benchmark-result.md"
}
},
"release_selfcheck": {
"what": "re-ran the 58 locked recordings through this package's own entry point and compared against the frozen prediction file",
"predictions_csv_sha256": "3ff22e6a55b13673b17d7d02ba5ac564ca777de96e927186172991508b144d69",
"byte_identical_to_frozen": true,
"max_abs_diff_abi_prime": 0.0,
"tolerance": 1e-06,
"status_rows_matching": "58/58 (57 scored / 1 rejected:R2|R5)",
"modules_verified_against_generation_manifest": "18/18 (13 stage 0 + 5 native) byte-identical",
"report": "aria-training/artifacts/release-selfcheck-v0.2.3.3/selfcheck_report.md"
},
"cold_start_thresholds": {
"T_diag_L0": 3.44,
"S_diag": 99.87,
"T_red_cold": 90.0,
"T_yellow_cold": 75.0,
"status": "UNCALIBRATED placeholder (T_diag from the L0 spec does not carry over to the severity label); the package therefore emits measurements only, no product band decision"
},
"supersedes": "v0.2.3.2 (draft, praat-script, archived in aria-training); v0.2.0.1 (breathiness label, archived in aria-training); v0.1-pre (this repo, archived under history/v0.1-pre/)"
}
|