Datasets:
name stringclasses 6
values | experiment_id int64 18 44 | lock_date timestamp[s]date 2026-02-18 00:00:00 2026-02-26 00:00:00 | overall_eer float64 3.13 4.7 | mean_attack_eer float64 1.91 3.09 | worst_eer float64 7.53 18.3 | worst_attack stringclasses 3
values | per_attack_eer dict | architecture stringclasses 6
values | codecs listlengths 3 3 | feature_types listlengths 5 8 | baseline_role stringclasses 1
value | canonical bool 1
class |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
9sys_+flux+f0 | 18 | 2026-02-18T00:00:00 | 4.44 | 3.09 | 14.63 | A18 | {
"A07": 1.11,
"A08": 0.34,
"A09": 0.02,
"A10": 2.3,
"A11": 0.42,
"A12": 1.27,
"A13": 0.14,
"A14": 5.55,
"A15": 2.36,
"A16": 1.17,
"A17": 11.48,
"A18": 14.63,
"A19": 0.38
} | 9 classifiers (3 codecs x 3 core features: LFCC/CQCC/MGDCC). Each classifier's input = core_feature(360-dim) + spectral_flux(12-dim) + f0_trajectory(24-dim) = 396-dim. Per-system StandardScaler, LogisticRegression(C=0.1, solver=lbfgs, max_iter=2000, seed=42). Min-max score normalization per system, avg_score fusion across 9 systems. | [
"wideband",
"g711_ulaw",
"amr_nb"
] | [
"lfcc",
"cqcc",
"mgdcc",
"spectral_flux",
"f0_trajectory"
] | canonical | true |
mega_all5 | 18 | 2026-02-18T00:00:00 | 4.7 | 3.04 | 18.31 | A14 | {
"A07": 0.78,
"A08": 0.26,
"A09": 0,
"A10": 1.85,
"A11": 0.32,
"A12": 0.95,
"A13": 0.09,
"A14": 18.31,
"A15": 1.99,
"A16": 0.8,
"A17": 5.73,
"A18": 8.06,
"A19": 0.42
} | 3 mega-classifiers (1 per codec). Each classifier's input = LFCC(360) + CQCC(360) + MGDCC(360) + spectral_flux(12) + f0_trajectory(24) = 1116-dim. Per-codec StandardScaler, LogisticRegression(C=0.1, solver=lbfgs, max_iter=2000, seed=42). Min-max score normalization per codec, avg_score fusion across 3 codecs. | [
"wideband",
"g711_ulaw",
"amr_nb"
] | [
"lfcc",
"cqcc",
"mgdcc",
"spectral_flux",
"f0_trajectory"
] | canonical | true |
delta_pareto_9sys | 20 | 2026-02-19T00:00:00 | 4.15 | 2.76 | 12.33 | A17 | {
"A07": 0.68,
"A08": 0.22,
"A09": 0.01,
"A10": 2.9,
"A11": 0.29,
"A12": 0.82,
"A13": 0.08,
"A14": 3.7,
"A15": 2.3,
"A16": 0.78,
"A17": 12.33,
"A18": 8.4,
"A19": 0.28
} | 9 classifiers (3 codecs x 3 core features: LFCC/CQCC/MGDCC). Each classifier's input = delta(static+Δ+ΔΔ, 1080-dim) + spectral_flux(12-dim) + f0_trajectory(24-dim) = 1116-dim. Per-system StandardScaler, LogisticRegression(C=0.1, solver=lbfgs, max_iter=2000, seed=42). Min-max score normalization per system, Pareto-weighted avg_score fusion (wideband=1.5, g711_ulaw=1.5, amr_nb=1.0). | [
"wideband",
"g711_ulaw",
"amr_nb"
] | [
"lfcc",
"cqcc",
"mgdcc",
"spectral_flux",
"f0_trajectory"
] | canonical | true |
exp37b_blend_1500d | 37 | 2026-02-22T00:00:00 | 3.13 | 1.91 | 9.88 | A17 | {
"A07": 0.56,
"A08": 0.7,
"A09": 0,
"A10": 1.45,
"A11": 0,
"A12": 0.5,
"A13": 0,
"A14": 1.78,
"A15": 1.82,
"A16": 0.7,
"A17": 9.88,
"A18": 6.05,
"A19": 1.45
} | 27 classifiers (3 codecs x 3 core features x 3 classifier types). Each classifier's input = delta(static+Δ+ΔΔ, 1080-dim) + spectral_flux(12-dim) + f0_trajectory(24-dim) + IFCC_delta(360-dim) + HF/LF_comod(24-dim) = 1500-dim. Three classifier types per (codec, core_feature): (1) Standard LR(C=0.1, solver=lbfgs, max_iter=2000, seed=42); (2) OHEM-LR: two-pass LR with loss-based sample weights from pass-1 (cap=5.0); (3) Shallow LGBM(n_estimators=200, max_depth=4, learning_rate=0.05, subsample=0.8, colsample_bytree=0.8, seed=42). Per-system StandardScaler, min-max score normalization. Pareto-weighted avg_score per classifier type (wideband=1.5, g711_ulaw=1.5, amr_nb=1.0). 3-component blend: score = 0.66 × OHEM + 0.04 × LGBM + 0.30 × LR. | [
"wideband",
"g711_ulaw",
"amr_nb"
] | [
"lfcc",
"cqcc",
"mgdcc",
"spectral_flux",
"f0_trajectory",
"ifcc",
"hf_lf_comod"
] | canonical | true |
exp40_cqcc_blend | 40 | 2026-02-24T00:00:00 | 3.35 | 2.4 | 7.53 | A17 | {
"A07": 1.46,
"A08": 1.47,
"A09": 0.01,
"A10": 2.45,
"A11": 0.24,
"A12": 0.1,
"A13": 0.01,
"A14": 2.41,
"A15": 3.06,
"A16": 1.37,
"A17": 7.53,
"A18": 6.62,
"A19": 4.48
} | 27 classifiers (3 codecs x 3 core features x 3 classifier types). Each classifier's input = delta(static+Δ+ΔΔ, 1080-dim) + spectral_flux(12-dim) + f0_trajectory(24-dim) + IFCC_delta(360-dim) + HF/LF_comod(24-dim) = 1500-dim. Three classifier types per (codec, core_feature): (1) Standard LR(C=0.1, solver=lbfgs, max_iter=2000, seed=42); (2) OHEM-LR: two-pass LR with loss-based sample weights from pass-1 (cap=5.0); (3) Shallow LGBM(n_estimators=200, max_depth=4, learning_rate=0.05, subsample=0.8, colsample_bytree=0.8, seed=42). Per-system StandardScaler, min-max score normalization. CQCC-weighted Pareto fusion (CQCC=1.5, LFCC=0.75, MGDCC=0.25) with codec weights (wideband=1.5, g711_ulaw=1.5, amr_nb=1.0). 3-component blend: score = 0.60 × OHEM + 0.16 × LGBM + 0.24 × LR. | [
"wideband",
"g711_ulaw",
"amr_nb"
] | [
"lfcc",
"cqcc",
"mgdcc",
"spectral_flux",
"f0_trajectory",
"ifcc",
"hf_lf_comod"
] | canonical | true |
exp44_scc_cqcc_blend | 44 | 2026-02-26T00:00:00 | 3.48 | 2.54 | 7.83 | A17 | {
"A07": 2.15,
"A08": 1.88,
"A09": 0.02,
"A10": 2.62,
"A11": 0.13,
"A12": 0.14,
"A13": 0.14,
"A14": 2.65,
"A15": 2.9,
"A16": 1.85,
"A17": 7.83,
"A18": 5.25,
"A19": 5.5
} | 27 classifiers (3 codecs x 3 core features x 3 classifier types). Each classifier's input = delta(static+Δ+ΔΔ, 1080-dim) + spectral_flux(12-dim) + f0_trajectory(24-dim) + IFCC_delta(360-dim) + HF/LF_comod(24-dim) + SCC_delta(360-dim) = 1860-dim. SCC = wavelet scattering transform (J=2, Q=10, order=2) → log → DCT (20 coeffs). Three classifier types per (codec, core_feature): (1) Standard LR(C=0.1, solver=lbfgs, max_iter=2000, seed=42); (2) OHEM-LR: two-pass LR with loss-based sample weights from pass-1 (cap=5.0); (3) Shallow LGBM(n_estimators=200, max_depth=4, learning_rate=0.05, subsample=0.8, colsample_bytree=0.8, seed=42). Per-system StandardScaler, Platt-calibrated scores (LogisticRegression on 1D). CQCC-weighted Pareto fusion (CQCC=2.5, LFCC=0.5, MGDCC=0.25, strong-B) with codec weights (wideband=1.5, g711_ulaw=1.5, amr_nb=1.0). 3-component blend: score = 0.47 × OHEM + 0.15 × LGBM + 0.38 × LR. | [
"wideband",
"g711_ulaw",
"amr_nb"
] | [
"lfcc",
"cqcc",
"mgdcc",
"spectral_flux",
"f0_trajectory",
"ifcc",
"hf_lf_comod",
"scc_delta"
] | canonical | true |
Sonotheia Calibration Baselines
Frozen calibration baselines from the Sonotheia voice fraud detection platform.
Overview
These baselines represent Sonotheia's interpretable, physics-based approach to voice spoofing detection evaluated on ASVspoof 2019 LA (eval partition, attacks A07–A19). No neural networks — linear classifiers on hand-crafted acoustic features with codec-aware calibration.
| Baseline | Overall EER | Mean Attack EER | Worst EER | Worst Attack | Role |
|---|---|---|---|---|---|
| 9sys_+flux+f0 | 4.44% | 3.09% | 14.63% | A18 | canonical |
| mega_all5 | 4.70% | 3.04% | 18.31% | A14 | canonical |
| delta_pareto_9sys | 4.15% | 2.76% | 12.33% | A17 | canonical |
| exp37b_blend_1500d | 3.13% | 1.91% | 9.88% | A17 | canonical |
| exp40_cqcc_blend | 3.35% | 2.40% | 7.53% | A17 | canonical |
| exp44_scc_cqcc_blend | 3.48% | 2.54% | 7.83% | A17 | canonical |
Best Configuration: exp37b_blend_1500d
- Overall EER: 3.13% on ASVspoof 2019 LA eval set
- 27 classifiers (3 codecs x 3 features x 3 classifier types)
- 1500-dim feature vector per classifier
- Pareto-weighted codec fusion + 3-component blend (OHEM-LR + LGBM + LR)
Per-Attack EER Breakdown
| Attack | 9sys_+flux+f0 | mega_all5 | delta_pareto_9sys | exp37b_blend_1500d | exp40_cqcc_blend | exp44_scc_cqcc_blend |
|---|---|---|---|---|---|---|
| A07 | 1.11% | 0.78% | 0.68% | 0.56% | 1.46% | 2.15% |
| A08 | 0.34% | 0.26% | 0.22% | 0.70% | 1.47% | 1.88% |
| A09 | 0.02% | 0.00% | 0.01% | 0.00% | 0.01% | 0.02% |
| A10 | 2.30% | 1.85% | 2.90% | 1.45% | 2.45% | 2.62% |
| A11 | 0.42% | 0.32% | 0.29% | 0.00% | 0.24% | 0.13% |
| A12 | 1.27% | 0.95% | 0.82% | 0.50% | 0.10% | 0.14% |
| A13 | 0.14% | 0.09% | 0.08% | 0.00% | 0.01% | 0.14% |
| A14 | 5.55% | 18.31% | 3.70% | 1.78% | 2.41% | 2.65% |
| A15 | 2.36% | 1.99% | 2.30% | 1.82% | 3.06% | 2.90% |
| A16 | 1.17% | 0.80% | 0.78% | 0.70% | 1.37% | 1.85% |
| A17 | 11.48% | 5.73% | 12.33% | 9.88% | 7.53% | 7.83% |
| A18 | 14.63% | 8.06% | 8.40% | 6.05% | 6.62% | 5.25% |
| A19 | 0.38% | 0.42% | 0.28% | 1.45% | 4.48% | 5.50% |
Codec Conditions
All baselines evaluate across three codec conditions:
- wideband: Clean 16kHz (no codec)
- g711_ulaw: G.711 mu-law (PSTN telephony)
- amr_nb: AMR-NB (mobile telephony)
Files
baselines.json— Full structured data with architecturesbaselines.csv— Flat CSV (one row per baseline x attack) for analysis
Citation
If you use these baselines for comparison, please cite:
Sonotheia: Interpretable Voice Fraud Detection with Codec-Aware Calibration
https://sonotheia.ai
License
Apache-2.0
- Downloads last month
- 2