README.md

metadata

pretty_name: ChainBench-ADD
license: other
license_name: chainbench-add-dataset-terms-of-use
tags:
  - audio
  - speech
  - deepfake-detection
  - benchmark
  - security
  - robustness
task_categories:
  - audio-classification
language:
  - en
  - zh

ChainBench-ADD

ChainBench-ADD is a delivery-aware benchmark for audio deepfake detection. It is designed to evaluate whether audio deepfake detectors remain reliable after realistic post-generation delivery effects such as compression, telephony degradation, replay-style simulation, and hybrid transformation chains.

Dataset Description

Dataset Summary

ChainBench-ADD focuses on post-generation delivery effects in audio deepfake detection. Rather than evaluating only clean bona fide and clean spoofed speech, it organizes samples into controlled delivery conditions so that robustness can be studied under more realistic downstream transformations.

The benchmark includes both bona fide and spoof speech, supports English and Mandarin Chinese, and provides metadata for structured evaluation across delivery conditions, operator chains, and protocol splits.

Dataset Composition

Labels

• bona_fide
• spoof

Scale

• Total waveforms: 941,201
• Bona fide samples: 315,573
• Spoof samples: 625,628
• Speakers: 448

Splits

The benchmark uses speaker-disjoint train/dev/test splits.

• Train: 663,363 samples from 314 speakers
• Dev: 136,027 samples from 67 speakers
• Test: 141,811 samples from 67 speakers

Sources

ChainBench-ADD is built from upstream speech resources and speech-generation systems, including:

• Bona fide sources: Common Voice 24.0, AISHELL-3
• Spoof generation systems: Qwen3-TTS, CosyVoice3, Spark-TTS, F5-TTS, IndexTTS2, VoxCPM

Metadata

Metadata include release and benchmarking information such as:

• sample identifiers
• parent identifiers
• split labels
• ground-truth labels
• language
• speaker identifiers
• transcript
• provenance/source information
• generator family and generator model
• delivery family
• template identifiers
• variant identifiers
• ordered operator signatures
• structural annotations required to instantiate benchmark tasks

Intended Use

ChainBench-ADD is released for:

• research on audio deepfake detection
• robustness evaluation
• forensic analysis
• provenance analysis
• benchmarking
• academic reproduction and verification

This dataset is intended for defensive, scientific, and evaluative use.

Out-of-Scope / Prohibited Use

ChainBench-ADD must not be used to support or enable:

• impersonation or identity fraud
• harassment, social engineering, or deception
• unauthorized voice cloning
• misleading, deceptive, or fraudulent media generation
• biometric surveillance or other high-risk deployment
• training, adapting, or improving systems intended for deceptive speech generation

Access and Redistribution

By accessing, downloading, or using this repository, you agree to the terms in the LICENSE file.

Redistribution of this benchmark package, or any subset that includes third-party material, is permitted only to the extent allowed by all applicable upstream terms and must preserve:

• the LICENSE file
• this dataset card
• all applicable upstream notices
• any required attribution and usage restrictions

Acknowledgements

ChainBench-ADD builds on upstream speech datasets and speech-generation systems. Users are responsible for consulting and complying with the original terms and attribution requirements of all upstream resources.