Datasets:

SpeechAntiSpoofingBenchmarks
/

CD-ADD

license: cc-by-4.0
language:
  - en
pretty_name: CD-ADD
task_categories:
  - audio-classification
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*.parquet
tags:
  - anti-spoofing
  - audio-deepfake-detection
  - speech
  - benchmark
  - arena-ready
paperswithcode_id: null
arxiv:
  - '2404.04904'

CD-ADD

Benchmark-ready packaging of CD-ADD (Cross-Domain Audio Deepfake Detection) for speech anti-spoofing / synthetic-voice detection.

Overview

CD-ADD pairs genuine speech with deepfakes produced by five advanced zero-shot text-to-speech systems (OpenVoice, VALL-E, WhisperSpeech, YourTTS, Seamless). This repo packages the LibriTTS test-clean subset together with all TED talks from CD-ADD into a single dataset. The task is binary classification: bonafide (genuine human speech) vs. spoof (zero-shot-TTS synthesized speech). Source paper: https://arxiv.org/abs/2404.04904.

License & redistribution

Redistributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license; the full text is in LICENSE.txt. Audio is the original 16 kHz mono signal losslessly re-encoded to FLAC (bit-exact samples); each real clip is labelled bonafide and every TTS-generated clip is labelled spoof.

Schema

Field	Type	Description
path	string	Source-relative path, unique within the dataset (e.g. `dataset_LibriTTS/test-clean/2300/131720/2300_131720_000002_000001/valle.wav`).
audio	Audio(16kHz mono)	Embedded 16 kHz mono audio.
label	ClassLabel[bonafide, spoof]	Index 0 = bonafide, 1 = spoof.
notes	string (JSON)	Contains a unique `utterance_id`, plus `source` (`libritts`/`ted`), `system` (`real`/`openvoice`/`valle`/`whisperSpeech`/`yourTTS`/`seamless`), `speaker_id`, and `transcript` when available.

Quick Start

from datasets import load_dataset
ds = load_dataset("SpeechAntiSpoofingBenchmarks/CD-ADD", split="test")

Stats

n_total	n_bonafide	n_spoof	total duration
20,786	3,661	17,125	~58.4 h

Source provenance

Built from the CD-ADD release (LibriTTS test-clean + all TED talks). Each source utterance directory contributes one real.wav (bonafide) and up to five zero-shot-TTS spoofs. LibriTTS dev-clean / train-clean-100 are excluded. Every decodable clip is included; clips that fail to decode via soundfile are dropped (none were dropped in this build).

Evaluation

See eval.yaml and submissions/README.md. Primary metric: EER (%), lower is better.

Citation

Original paper: https://arxiv.org/abs/2404.04904

@article{li2024crossdomain,
  title   = {Cross-Domain Audio Deepfake Detection: Dataset and Analysis},
  author  = {Li, Yuang and Zhang, Min and Ren, Mengxin and Ma, Miaomiao and
             Wei, Daimeng and Yang, Hao},
  journal = {arXiv preprint arXiv:2404.04904},
  year    = {2024}
}

Maintainer

SpeechAntiSpoofingBenchmarks — contact k.n.borodin@mtuci.ru