SpeechAntiSpoofingBenchmarks/ASVspoof2019_LA

ASVspoof2019LA

ASVspoof2019LA is a benchmark-ready packaging of the Logical Access (LA) evaluation partition from ASVspoof 2019 for speech spoofing and synthetic voice detection research.

This repository is intended for Hugging Face hosting and Papers with Code style benchmark tracking. It contains only the LA evaluation audio and the official LA countermeasure evaluation protocol. The original dataset at was not modified.

Contact: k.n.borodin@mtuci.ru

Contents

ASVspoof2019LA/
  data/
    ASVspoof2019LA_eval_flac.tar.gz
  protocols/
    ASVspoof2019.LA.cm.eval.trl.txt
  submissions/
    README.md
    results_template.csv
  LICENSE.txt
  README.md

The archive data/ASVspoof2019LA_eval_flac.tar.gz extracts to:

flac/
  LA_E_*.flac

Benchmark Task

Given an evaluation utterance from the ASVspoof 2019 Logical Access condition, predict whether it is:

• bonafide: genuine human speech
• spoof: synthetic or converted speech

The benchmark uses the official ASVspoof 2019 LA countermeasure evaluation protocol:

protocols/ASVspoof2019.LA.cm.eval.trl.txt

Protocol columns:

speaker_id utterance_id unused attack_id label

Example:

LA_0039 LA_E_2834763 - A11 spoof
LA_0004 LA_E_1665632 - - bonafide

Number of evaluation trials: 71,237.

Recommended Metrics

Report both metrics when possible:

• Equal Error Rate (EER), lower is better.
• Minimum tandem Detection Cost Function (min t-DCF), lower is better.

Use the official ASVspoof 2019 scoring implementation when comparing with published challenge results.

How to Use

Download and extract the evaluation audio:

tar -xzf data/ASVspoof2019LA_eval_flac.tar.gz -C data/

This creates data/flac/*.flac. Match each utterance_id in the protocol to data/flac/<utterance_id>.flac.

Adding Results

To add benchmark results:

1. Evaluate your system on every trial in protocols/ASVspoof2019.LA.cm.eval.trl.txt.
2. Compute EER and min t-DCF using the official ASVspoof 2019 evaluation code.
3. Add your result to submissions/results_template.csv or submit a new CSV file under submissions/.
4. Include links to the paper, code, score file, and contact email when available.
5. Open a pull request or contact k.n.borodin@mtuci.ru.

For Papers with Code visibility, include the same benchmark name, dataset name, metrics, paper link, and code link in your paper/code metadata. Use:

• Dataset: ASVspoof2019LA
• Task: Speech Spoofing Detection
• Benchmark split: ASVspoof 2019 LA evaluation
• Metrics: EER (%) and min t-DCF

Hugging Face Upload Notes

Large files are configured for Git LFS via .gitattributes.

Suggested upload workflow:

cd ASVspoof2019LA
git init
git lfs install
git add .gitattributes README.md LICENSE.txt protocols submissions data/ASVspoof2019LA_eval_flac.tar.gz
git commit -m "Add ASVspoof2019LA benchmark dataset"
git remote add origin https://huggingface.co/datasets/<user-or-org>/ASVspoof2019LA
git push -u origin main

Citation

If you use this benchmark, cite the original ASVspoof 2019 database paper:

@article{wang2020asvspoof,
  title={ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech},
  author={Wang, Xin and Yamagishi, Junichi and Todisco, Massimiliano and Delgado, Hector and Nautsch, Andreas and Evans, Nicholas and Sahidullah, Md and Vestman, Ville and Kinnunen, Tomi and Lee, Kong Aik and others},
  journal={Computer Speech & Language},
  volume={64},
  pages={101114},
  year={2020},
  publisher={Elsevier}
}

License

This benchmark package includes LICENSE.txt copied from the original ASVspoof 2019 LA distribution. Users are responsible for following the original dataset license and citation requirements.