Datasets:
YAML Metadata Error:Invalid content in eval.yaml.
Check out the documentation for more information.
Show details
✖ Invalid input
Download eval.yaml from SpeechAntiSpoofingBenchmarks/CFAD: direct link, hf CLI and curl.
- Browser
- Download file 709 Bytes
-
https://huggingface.co/datasets/SpeechAntiSpoofingBenchmarks/CFAD/resolve/refs%2Fpr%2F13/eval.yaml
- Command line
-
hf download hf://datasets/SpeechAntiSpoofingBenchmarks/CFAD@refs/pr/13/eval.yaml
-
curl -L -o eval.yaml https://huggingface.co/datasets/SpeechAntiSpoofingBenchmarks/CFAD/resolve/refs%2Fpr%2F13/eval.yaml
709 Bytes
| 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 | |