QuadVoxBench: A Large-Scale Fine-Grained Benchmark for Robust Audio Deepfake Detection

Dataset Description

QuadVoxBench is a large-scale (574.5+ hours) audio benchmark dataset designed for the robust evaluation of deepfake detection systems. It is structured around four key aspects of audio variation: Speech Style, Emotional Prosody, Acoustic Environment, and Manipulation Type. The dataset features a diverse collection of real and synthetically generated audio in English and Chinese, created using a comprehensive toolkit of modern text-to-speech (TTS) and voice conversion (VC) models.

For more details, please refer to our main repository:

Dataset Structure

The dataset is organized into 11 subsets, each corresponding to a specific domain or speech characteristic. Each subset contains real and fake audio samples, along with metadata.

├── Audiobook/
│   ├── audio/
│   │   ├── real/
│   │   └── fake/
│   └── meta_test.json
├── Emotional/
│   ├── audio/
│   │   ├── real/
│   │   └── fake/
│   └── meta_test.json
├── Interview/
│   ├── audio/
│   │   ├── real/
│   │   └── fake/
│   └── meta_test.json
├── Movie/
│   ├── audio/
│   │   ├── real/
│   │   └── fake/
│   └── meta_test.json
├── News/
│   ├── audio/
│   │   ├── real/
│   │   └── fake/
│   └── meta_test.json
├── NoisySpeech/
│   ├── audio/
│   │   ├── real/
│   │   └── fake/
│   └── meta_test.json
├── PartialFake/
│   ├── audio/
│   │   ├── real/
│   │   └── fake/
│   └── meta.json
├── PhoneCall/
│   ├── en/
│   │   ├── audio/
│   │   │   ├── real/
│   │   │   └── fake/
│   │   └── meta_test.json
│   └── zh-cn/
│       ├── audio/
│       │   ├── real/
│       │   └── fake/
│       └── meta_test.json
├── Podcast/
│   ├── audio/
│   │   ├── real/
│   │   └── fake/
│   └── meta_test.json
├── PublicFigure/
│   ├── audio/
│   │   ├── real/
│   │   └── fake/
│   └── meta_test.json
└── PublicSpeech/
    ├── audio/
    │   ├── real/
    │   └── fake/
    └── meta_test.json