name: DFADD
description: >
  DFADD (The Diffusion and Flow-Matching Based Audio Deepfake Dataset, arXiv
  2409.08731) — VCTK-derived anti-spoofing set targeting modern diffusion /
  flow-matching TTS. Eval (test) split only. Binary classification: bonafide
  (genuine VCTK) vs. spoof (Grad-TTS / Matcha-TTS / NaturalSpeech 2 / PFlow-TTS
  / StyleTTS 2). EER over the full test set (3,755 utterances: 755 bonafide /
  3,000 spoof).
evaluation_framework: inspect-ai

tasks:
  - id: antispoofing_eval
    config: default
    split: test

    field_spec:
      input: audio
      target: label

    solvers:
      - name: speech_spoof_bench_solver

    scorers:
      - name: speech_spoof_scorer

    metrics:
      - eer_percent