name: LRLspoof
description: >
  LRLspoof - a large multilingual low-resource-language anti-spoofing benchmark
  of 1,304,455 text-to-speech (spoof) utterances spanning 66 languages and many
  TTS systems. It is SPOOF-ONLY (no bonafide), so it is scored with the 1-SRR
  metric (srr_complement): the fraction of spoof NOT rejected at a fixed
  operating threshold t* transferred from DeepVoice (lower is better). Every
  model must carry a DeepVoice submission to derive t* (placed in the
  submission's `calibration` block). The 500+ GB audio is NOT stored in this
  arena entry - it lives as a multi-part tarball in the source repo; only
  data/labels.parquet (utterance_id -> label, every label = 1 = spoof) is
  shipped here. utterance_id = the audio file's path relative to the dataset
  root, e.g. "english/fastpitch/line_59.wav".
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:
      - srr_complement