2024
DOI: 10.48550/arxiv.2401.09512
| |Summarize |Cite
Preprint
|
Sign up to set email alerts

MLAAD: The Multi-Language Audio Anti-Spoofing Dataset

Search citation statements

Order By: Relevance

Paper Sections

Select...
7
2
1
0

Citation Types

0
9
0
0

Year Published

Range
2024
2024
2026
2026

Publication Types

Select...
3
3
1

Relationship

0
7

Authors

Journals

Cited by 6 publications

(9 citation statements)

References 0 publications

0
9
0
0
Order By: Relevance
How the paper below cites the one you are viewing
“…Datasets: ASVspoof 2019-LA (Todisco et al, 2019), 19 TTS/VC spoofing methods under controlled conditions; FakeOrReal (FoR) (Reimao & Tzerpos, 2020), 8 TTS systems, 2 s clips, English; InTheWild (ITW) (Müller et al, 2022), celebrity deepfakes with real-world noise and compression; MLAAD v1 (Müller et al, 2024), 54 TTS systems across 23 languages (the most comprehensive benchmark). Zero-shot transfer is additionally evaluated on WaveFake (Frank & Schönherr, 2021) (spoof-only) and LJSpeech (Ito, 2017) (bonafide-only).…”
Section: Methods
mentioning
confidence: 99%
“…The training set remains completely unseen during evaluation without target-domain adaptation, mirroring unseen practical deployment conditions. Baselines (RawGAT-ST, SSL W2V2, Whisper-DF) are from Müller et al (2024). We report Accuracy (ACC) and Equal Error Rate (EER).…”
Section: Methods
mentioning
confidence: 99%
“…We run 7 independent seeds (42-48) controlling weight initialisation and data-shuffle order, and retain the top-5 by held-out accuracy. Seeds are fixed before any experiment and are never tuned; we treat the seed as a reproducibility hyperparameter, matching the MLAAD benchmark protocol (Müller et al, 2024).…”
Section: Conclusion and Responsible Deployment
mentioning
confidence: 99%
“…"-": in-distribution, excluded per protocol. Baselines fromMüller et al (2024). WaveFake (spoof-only) and LJSpeech (bonafide-only) are zero-shot held-out sets.…”
mentioning
confidence: 99%
See 3 more Smart Citations