AuralGuard: Multi-View One-Class Learning for Generalizable and Calibrated Detection of AI-Generated Speech in the Wild
Official checkpoint repository and benchmark artifacts for the paper:
"AuralGuard: Multi-View One-Class Learning for Generalizable and Calibrated Detection of AI-Generated Speech in the Wild"
Target: Elsevier Computer Speech & Language / IEEE Transactions on Information Forensics and Security (IEEE TIFS)
- 📄 Manuscript (18 pages): not currently hosted in this repository — the former
paper/main.pdffile was removed from the repo (rev0147823630); contact the authors for a copy. - 💻 Official Code Repository: github.com/MIHMahmudEli/auralguardv2
- 👥 Authors: Mohsin Ibna Hossain, Md. Muttakin, Sadman Hasan Miraj, and Mohammad Mahmudul Hasan (Corresponding Author)
- 🏛️ Affiliation: Department of Computer Science, American International University-Bangladesh (AIUB)
📊 Empirical Benchmark Results
Trained strictly once on ASVspoof 2019 LA and evaluated zero-shot out-of-domain across all benchmarks. Metrics report point estimates with 95% non-parametric bootstrap confidence intervals ($B=1,000$).
| Architecture | In-Domain (19LA) EER (%) | min t-DCF | Zero-Shot (In-the-Wild) EER (%) | ITW AUROC | ITW ECE (↓) | Zero-Shot (WaveFake) EER (%) |
|---|---|---|---|---|---|---|
| B1 (LFCC-LCNN) | 16.29 [15.97, 16.63] | 0.6131 | 50.62 [50.00, 51.19] | 0.4979 | 0.4292 | 45.32 [44.86, 45.76] |
| B2 (RawNet2) | 8.99 [8.72, 9.23] | 0.3444 | 38.08 [37.57, 38.59] | 0.6671 | 0.5761 | 50.37 [49.89, 50.80] |
| B3 (AASIST) | 10.63 [10.36, 10.92] | 0.3921 | 42.33 [41.80, 42.85] | 0.6122 | 0.5552 | 49.89 [49.37, 50.34] |
| B5 (WavLM+OCS) | 2.86 [2.73, 3.00] | 0.1041 | 16.55 [16.12, 16.98] | 0.9103 | 0.2320 | 47.43 [46.91, 47.89] |
| AuralGuard (v1) | 3.58 [3.40, 3.74] | 0.1326 | 26.02 [25.55, 26.52] | 0.8145 | 0.1028 | 30.66 [30.27, 31.02] |
| AuralGuard v2 | 6.81 [6.61, 7.00] | 0.2398 | 19.19 [18.78, 19.61] | 0.8820 | 0.1716 | 45.19 [44.69, 45.71] |
Provenance note (2026-10-22, independent pre-submission audit): all non-v1 rows are computed from this repository's
results/*.json— their point estimates were already consistent, but several displayed confidence intervals had drifted from the JSONs and are now regenerated directly from them. The AuralGuard (v1) row is an independent re-evaluation of the releasedcheckpoints/auralguard/best.ckpt(sha25684cb085b0d4dab59617b7090de2782fa07650e59252e7f0f641fade1695ba171, epoch 16) using this repository's own evaluation code and protocol (bootstrap $B=1{,}000$): in-domain 3.58 %, In-the-Wild 26.02 %, WaveFake 30.66 %. The v1 result JSONs in this repository predate the final checkpoint upload (3.42 / 24.11 / 49.86 %) and the previously displayed README values (7.30 / 23.47 / 48.72 %) are superseded — neither reproduces from the released checkpoint.
📈 Actual Manuscript Figures
1. Cross-Domain Generalization (Fig. 6)
2. Detection Error Tradeoff (DET) Curves (Fig. 7)
3. Non-Parametric 95% Bootstrap Confidence Intervals (Fig. 8)
4. Discrimination vs. Probability Calibration (Fig. 9)
5. Out-of-Domain Reliability Diagram on In-the-Wild (Fig. 10)
📦 Checkpoint Directory
| Model | Path in Repo | Description |
|---|---|---|
| AuralGuard v2 | checkpoints/auralguard_v2/best.ckpt |
Multi-view + Gated Cross-Attention + AASIST + OCS + SupCon |
| AuralGuard (v1) | checkpoints/auralguard/best.ckpt |
Multi-view + AASIST + OCS |
| B5 (WavLM+OCS) | checkpoints/b5_wavlm_ocs/best.ckpt |
Single-view WavLM-Large + AASIST + OCS |
| B3 (AASIST) | checkpoints/b3_aasist/best.ckpt |
Sinc-conv raw waveform + Graph attention network |
| B2 (RawNet2) | checkpoints/b2_rawnet2/best.ckpt |
Sinc-conv raw waveform + residual blocks + GRU |
| B1 (LFCC-LCNN) | checkpoints/b1_lcnn/best.ckpt |
60-dim LFCC + Max-Feature-Map CNN |
🚀 Quick Usage
from huggingface_hub import hf_hub_download
import torch
# Download AuralGuard v2 checkpoint
ckpt_path = hf_hub_download(
repo_id="MoshinAli/auralguard-checkpoints",
filename="checkpoints/auralguard_v2/best.ckpt"
)
print(f"Downloaded checkpoint to: {ckpt_path}")