aria_linear_head
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/)"
}