eval.yaml

name: CVoiceFake_small
description: >
  CVoiceFake (small) multilingual audio-deepfake benchmark from the SafeEar paper
  (arXiv 2409.09272). The public 10% subset across five languages (de, en, fr, it,
  zh-CN), sourced from Mozilla CommonVoice. Binary classification: bonafide
  (genuine CommonVoice speech) vs. spoof (re-synthesised by five vocoders: Griffin-
  Lim, WORLD, Parallel WaveGAN, Multi-band MelGAN, Style MelGAN). EER over all
  138136 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