qhjqhj00 speech-deepfake-detection-eval
ProSDD: Learning Prosodic Representations for Speech Deepfake Detection against Expressive and Emotional Attacks — Mahapatra et al. (2026) (arXiv:2604.13229, 2026)
What this evaluates
Evaluates a model's ability to distinguish between authentic human speech and synthetically generated or manipulated speech (deepfakes), with a specific focus on robustness against expressive and emotional synthesis attacks.
Datasets
- LibriSpeech — total ?; splits: train (-1), dev (-1)
- ASVspoof 2019 LA — total ?; splits: train (-1), dev (-1), eval (-1)
- ASVspoof 2021 LA — total ?; splits: eval (-1)
- ASVspoof 2024 — total ?; splits: train (-1), dev (-1), eval (-1)
- EmoFake — total ?; splits: eval (-1)
- EmoSpoof-TTS — total ?; splits: eval (-1)
Metrics
EER (Equal Error Rate)(primary) — range: percent- Standard spoof detection metric; not explicitly named in this section but universally used for ASVspoof and EmoFake benchmarks. It is the operating point where the false acceptance rate equals the false rejection rate.
validation accuracy— range: percent- Percentage of correctly classified 4-second audio segments on the development set, used for model selection during training.
Input / output format
Input: Fixed 4-second audio segments containing either bona fide or spoofed speech.
Output: Binary classification prediction (bona fide vs. spoofed) or spoof probability score.
Scoring recipe
def compute_eer(scores, labels):
# scores: model output probabilities for spoof class
# labels: ground truth (1=spoof, 0=bona fide)
far_list, frr_list = [], []
for threshold in np.linspace(0, 1, 1000):
far = sum(1 for s, l in zip(scores, labels) if s >= threshold and l == 0) / max(sum(1 for l in labels if l == 0), 1)
frr = sum(1 for s, l in zip(scores, labels) if s < threshold and l == 1) / max(sum(1 for l in labels if l == 1), 1)
far_list.append(far)
frr_list.append(frr)
# Find threshold where FAR ≈ FRR
eer = np.interp(0.5, far_list, frr_list)
return eer * 100
Common pitfalls
- Using the development set for final metric reporting instead of the official evaluation set.
- Training baselines inconsistently (e.g., using official pretrained checkpoints for ASVspoof 2019 but training from scratch for ASVspoof 2024).
- Ignoring the official train/dev/eval splits defined by each benchmark organization.
Evidence (verbatim from paper)
For traditional benchmarks, we evaluate on ASVspoof 2019 LA and ASVspoof 2021 LA. For emotional and expressive evaluation, we use EmoFake, EmoSpoof-TTS (abbreviated as EmoSpoof in Tables), and ASVspoof 2024 Track 1, which includes modern expressive synthesis systems. All experiments follow the official train/dev/eval splits of each dataset.
Citation
@misc{mahapatra2026prosdd,
title={ProSDD: Learning Prosodic Representations for Speech Deepfake Detection against Expressive and Emotional Attacks},
author={Mahapatra et al. (2026)},
year={2026},
note={arXiv:2604.13229}
}
- arXiv: 2604.13229