eval.yaml

name: CD-ADD
description: >
  Cross-Domain Audio Deepfake Detection (CD-ADD). The LibriTTS test-clean subset
  and all TED talks from CD-ADD, packaged as a single benchmark. Each genuine
  utterance is paired with zero-shot-TTS spoofs from up to five systems
  (OpenVoice, VALL-E, WhisperSpeech, YourTTS, Seamless). Binary classification:
  bonafide vs. spoof, scored by EER.
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