Datasets:
File size: 1,134 Bytes
5caf113 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 | 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
|