SpeechAntiSpoofingBenchmarks/CVoiceFake_small

CVoiceFake (small)

Benchmark-ready packaging of the CVoiceFake (small) multilingual audio-deepfake detection set introduced with SafeEar (arXiv 2409.09272), for speech anti-spoofing / synthetic-voice detection.

Overview

CVoiceFake is a multilingual deepfake-audio dataset built on top of Mozilla CommonVoice: genuine clips are re-synthesised by a bank of vocoders to produce matched spoof audio. This repo packages the public small subset — a random ~10% of the full collection — across five languages (German, English, French, Italian, Chinese). It is a binary classification benchmark: bonafide (genuine CommonVoice speech) vs. spoof (vocoder re-synthesis).

The spoof side aggregates five synthesis systems, each applied to the same bonafide utterances:

• Griffin-Lim (classical phase reconstruction)
• WORLD (vocoder)
• Parallel WaveGAN
• Multi-band MelGAN
• Style MelGAN

(The SafeEar paper additionally describes a DiffWave system; it is not present in this public small subset, which ships the five systems above.)

License & redistribution

Released under CC BY 4.0 (Creative Commons Attribution 4.0 International), the upstream license on the CVoiceFake Zenodo record — see LICENSE.txt (verbatim) and RIGHTS.md for provenance. CC BY 4.0 permits redistribution and derivatives (including the format normalisation done here) with attribution, so this rehost is legally clean. Underlying speech is from Mozilla CommonVoice.

Schema

Field	Type	Description
path	string	Source-relative path (<lang>/<system>/<file>.mp3), unique.
audio	Audio(16kHz mono)	Source MP3 bytes embedded verbatim (already 16 kHz mono).
label	ClassLabel[bonafide, spoof]	Index 0 = bonafide, 1 = spoof.
notes	string (JSON)	utterance_id, system_id, language, source.

notes example (spoof): {"utterance_id": "CVF_de_world_generated_common_voice_de_38024994_Gen", "system_id": "world_generated", "language": "de", "source": "CommonVoice"}

Quick Start

from datasets import load_dataset
ds = load_dataset("SpeechAntiSpoofingBenchmarks/CVoiceFake_small", split="test")

Stats

n_total	n_bonafide	n_spoof	total duration
138,136	23,544	114,592	~218 hours

Per-language (bonafide / spoof):

Lang	Bonafide	Spoof	Total
de	7,251	36,064	43,315
en	4,315	21,438	25,753
fr	3,589	17,685	21,274
it	3,865	18,052	21,917
zh-CN	4,524	21,353	25,877

All clips are 16 kHz mono. The spoof side dominates because each bonafide clip is re-synthesised by up to five systems.

Source provenance

• Bonafide: genuine speech from Mozilla CommonVoice (per language).
• Spoof: vocoder re-synthesis of the bonafide clips (Griffin-Lim, WORLD, Parallel WaveGAN, Multi-band MelGAN, Style MelGAN).
• Upstream: CVoiceFake / SafeEar — Zenodo record 11124319; project page https://safeearweb.github.io/Project/.

Evaluation

See eval.yaml and submissions/README.md. Primary metric: EER (%) over all 138,136 utterances.

Citation

Original paper: SafeEar: Content Privacy-Preserving Audio Deepfake Detection (arXiv 2409.09272).

@inproceedings{li2024safeear,
  title     = {SafeEar: Content Privacy-Preserving Audio Deepfake Detection},
  author    = {Li, Xinfeng and Li, Kai and Zheng, Yifan and Yan, Chen and Ji, Xiaoyu and Xu, Wenyuan},
  booktitle = {Proceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security (CCS)},
  year      = {2024},
  doi       = {10.1145/3658644.3670285}
}

Maintainer

Packaged for the Speech Anti-Spoofing Arena (SpeechAntiSpoofingBenchmarks). Maintained by Kirill Borodin (SpeechAntiSpoofingBenchmarks).

• Email: k.n.borodin@mtuci.ru (deprecated — use kborodin.research@gmail.com)
• Telegram: @korallll_ai