Datasets:

Modalities:
Audio
Text
Formats:
parquet
Languages:
English
ArXiv:
License:

YAML Metadata Error:Invalid content in eval.yaml.

Check out the documentation for more information.

Show details
✖ Invalid option: expected one of "prompt_template"|"system_message"|"user_message"|"chain_of_thought"|"use_tools"|"generate"|"self_critique"|"multiple_choice" → at tasks[0].solvers[0].name ✖ Invalid option: expected one of "includes"|"match"|"pattern"|"answer"|"exact"|"f1"|"model_graded_qa"|"model_graded_fact"|"choice" → at tasks[0].scorers[0].name
EmoSpoofTTS / eval.yaml
korallll's picture
Make arena-suitable: spoof-only -> srr_complement (1-SRR) metric + DeepVoice threshold-transfer; README/eval.yaml updated
0ee6c4d verified
Raw
History Blame Contribute Delete
984 Bytes
name: EmoSpoofTTS
description: >
Spoof-only attack corpus of emotional TTS synthesis (Mahapatra et al.,
"Can Emotion Fool Anti-spoofing?", Interspeech 2025): 36,000 clips from
3 TTS systems (StyleTTS2, F5-TTS, CosyVoice) x 10 ESD speakers x 4 emotions.
It contains NO bonafide audio (every label = 1 = spoof), so it is scored with
the 1-SRR metric (srr_complement): the fraction of these emotional-TTS spoofs
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). EER on this dataset alone is
degenerate and is not used.
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