AffectDF/AffectDF_EmotionSDD

AffectDF: Emotionally Expressive Speech Deepfake Benchmark

Overview

AffectDF is a large-scale benchmark for speech deepfake detection under emotionally expressive spoofing conditions. The dataset is designed to evaluate whether current speech deepfake detection (SDD) systems can generalize beyond conventional neutral-speech benchmarks to modern emotional and expressive speech attacks.

AffectDF contains approximately 260 hours of audio generated using 21 spoofing attacks across five emotional states. The dataset includes both acted emotional speech and spontaneous emotional speech, enabling analysis of robustness across different emotional speaking styles.

The dataset is organized into three official splits:

• Train/
• Dev/
• Test/

Protocol files are provided for reproducible evaluation.

---

Source Corpora

AffectDF is constructed using bona fide emotional speech from two source corpora:

• ESD (Emotional Speech Dataset): acted emotional speech
• MSP-Podcast: spontaneous emotional speech

The use of both corpora enables evaluation across controlled acted emotional speech and naturalistic spontaneous emotional speech.

---

Emotion Categories

AffectDF covers five emotional states:

• Angry
• Happy
• Neutral
• Sad
• Surprise

---

Dataset Design

AffectDF contains both bona fide and synthetic emotional speech samples.

• Bona fide samples are taken from ESD and MSP-Podcast.
• Synthetic samples are generated using diverse spoofing systems spanning TTS, VC, EVC, and LALM-based attacks.
• The dataset supports analysis across emotion labels, attack types, generation models, speakers, and acted versus spontaneous emotional speech.
• Many samples follow a parallel or controlled design, where synthetic samples can be aligned with corresponding source utterances using audio IDs and protocol metadata.

---

Spoofing Attack Types

AffectDF includes attacks from multiple generation paradigms:

• TTS: text-to-speech attacks
• VC: voice conversion attacks
• EVC: emotional voice conversion attacks
• VC+EVC: combined voice and emotional voice conversion attacks
• LALM-EVC: large audio-language model based emotional voice conversion attacks

The benchmark includes both conventional generation systems and recent large audio-language model based attacks.

---

Directory Structure

The dataset is organized as follows:

AffectDF/
├── Train/
├── Dev/
├── Test/
└── Protocols/
    ├── train_protocol.txt
    ├── dev_protocol.txt
    └── test_protocol.txt

Example audio filenames:

R_0015_000172.wav
CV_0016_001189.wav
A2S_0015_0015_000654_Angry_to_Sad.wav
MCPM_A2N_0017_0017_000613_Angry_to_Neutral.wav
StA_N2A_0015_000197.wav

---

Protocol Files

The Protocols/ directory contains split-specific metadata files:

Protocols/train_protocol.txt
Protocols/dev_protocol.txt
Protocols/test_protocol.txt

Each protocol entry contains metadata needed for evaluation and analysis.

Example entry:

0011 A2A_0011_0011_000351_Angry_to_Angry A01 Angry EVC Qwen2.5-Omni spoof

The fields include:

speaker_id audio_id attack_id emotion generation_type generation_model label

Where:

• speaker_id: speaker identifier
• audio_id: utterance or file identifier
• attack_id: attack system identifier
• emotion: emotion label
• generation_type: attack category, such as TTS, VC, EVC, or LALM-EVC
• generation_model: model used to generate the sample
• label: real or spoof

---

Bona Fide and Synthetic Pairing

AffectDF supports controlled comparison between bona fide and synthetic speech. Synthetic samples can be linked to corresponding or related bona fide samples through shared audio ID components in the filenames and protocol entries.

This design enables:

• real-versus-spoof comparison
• emotion-wise analysis
• attack-wise analysis
• acted versus spontaneous comparison

---

Ethical Considerations

In this work, we use publicly available datasets and models to generate synthetic speech for the purpose of improving speech deepfake detection. We do not release any additional speaker information beyond what is already available in the original source datasets, and we do not synthesize samples containing harmful or sensitive content. The generated data and analyses are intended solely to support research on robust detection, security, and responsible benchmarking of speech deepfakes. We do not intend for AffectDF or any generation pipeline described in this work to be used for impersonation, deception, or other malicious applications.

---

Data License and Intended Use

Data License and Intended Use. AffectDF is released for research and non-commercial use only under a CC BY-NC 4.0 license. The release includes generated audio and protocol files, subject to the licenses and terms of the original source corpora and generation models. The dataset is intended only for speech deepfake detection research and must not be used for impersonation, deception, biometric spoofing, or other harmful applications.