eval.yaml

name: CFAD
description: >
  CFAD (Chinese Fake Audio Detection, arXiv 2207.12308) clean test set. Binary
  anti-spoofing on Mandarin Chinese speech: 20,999 bonafide (aishell1/aishell3/
  magicread/thchs30/magicconversa/selfrecording) vs. 42,000 spoof from twelve
  synthesis/vocoder systems, split into test_seen (in-training systems) and
  test_unseen (held-out systems). EER computed over all 62,999 utterances.
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