Speech Anti-Spoofing Benchmarks

An open, reproducible benchmark for speech deepfake and anti-spoofing detection.

We collect public anti-spoofing test sets into one consistent format, score every detector on the same data, and publish the results on the Speech Anti-Spoofing Arena leaderboard. Every score on the board is independently reproduced from the submitted per-utterance scores before it is listed — no self-reported numbers.

What's here

• 🏟️ Arena leaderboard: 30+ detectors, from classic AASIST/RawNet-style models to large SSL-based systems.
• 📚 24 Core datasets, all packaged with the same layout and pinned revisions: ASVspoof 2019 LA, 2021 LA, 2021 DF, ASVspoof 5 · In-the-Wild · ADD 2022/2023 · CFAD · CD-ADD · CVoiceFake · DeepVoice · LibriSeVoc · SONAR · ArAD · DECRO · J-SPAW · ODSS · HABLA · DFADD · PyAra · XMAD · EmoFake · EmoSpoofTTS · LRLspoof
• 📋 arena-manifest: the single source of truth for which datasets and revisions count.

How ranking works

The main metric is EER (Equal Error Rate, %), where lower is better. Spoof-only sets use the share of fakes accepted at the model's calibrated threshold. Models are grouped into tiers by how many Core datasets they cover: 🥇 Gold (all), 🥈 Silver (at least half), 🥉 Bronze (at least one). Models without a paper are welcome too, and are listed in the 🔓 Unpublished / Proprietary tier. → How ranking works

Contribute

Everything goes through the open-source toolkit speech-spoof-bench:

• Submit a model: run the benchmark and open a PR on each dataset repo. CI re-scores it automatically. → Guide
• Add a dataset: → Guide
• Can't run the full benchmark yourself? Email us and we'll consider running your model for you.

Maintained by Lab260.

Contact

• 📧 Email: kborodin.research@gmail.com
• 💬 Telegram: @korallll_ai
• 🐙 GitHub: lab260ru/speech_spoof_bench (issues and PRs welcome)