🎙️ FaEmoSpoof

A Paired Persian Emotional Speech Deepfake Dataset

FaEmoSpoof is a Persian speech dataset for audio deepfake detection, anti-spoofing, and emotion-aware synthetic speech analysis.

It contains 2,848 matched Real/Fake pairs generated using OmniVoice zero-shot voice cloning.
Each pair shares the same speaker, transcript, and emotion label, enabling controlled comparison between authentic and synthetic Persian speech.

📊 Dataset at a Glance

Language Persian / Farsi	Real/Fake Pairs 2,848	Total Audio 5,696	Speakers 82
Real Audio 2,848	Synthetic Audio 2,848	Emotions 6	Generator OmniVoice

🎭 Emotion Distribution

Emotion	Matched Pairs	Audio Samples
😠 Anger	1,024	2,048
😐 Neutral	1,139	2,278
😢 Sadness	339	678
🙂 Happiness	180	360
😲 Surprise	149	298
😨 Fear	17	34
Total	2,848	5,696

🔗 Paired Design

For every synthetic utterance:

• speaker identity is matched
• target transcript is identical
• emotion label is matched
• cloning references come from the same speaker and emotion
• the target utterance itself is excluded from the cloning reference

This design enables Real/Fake comparison while controlling for speaker, linguistic content, and emotion label.

📁 Dataset Structure

FaEmoSpoof/
├── audio/
│   ├── real/
│   │   ├── anger/
│   │   ├── fear/
│   │   ├── happiness/
│   │   ├── neutral/
│   │   ├── sadness/
│   │   └── surprise/
│   └── fake/
│       ├── anger/
│       ├── fear/
│       ├── happiness/
│       ├── neutral/
│       ├── sadness/
│       └── surprise/
├── metadata.csv
├── pairs.csv
└── manifests/
    └── paired_metadata.csv

🗂️ Metadata

metadata.csv contains one row per audio sample (5,696 rows).

file_name · pair_id · label · label_id · speaker_id · gender · emotion · transcript · generator · language

pairs.csv contains one row for each of the 2,848 matched Real/Fake pairs.

🚀 Quick Start

Install:

pip install datasets

Load:

from datasets import load_dataset

dataset = load_dataset(
    "mobina1380/FaEmoSpoof",
    split="train"
)

print(dataset)
print(dataset[0])

Expected size:

5696 samples

Use:

• label → Real/Fake classification
• emotion → emotion-aware analysis
• pair_id → matched Real/Fake comparison
• speaker_id → speaker-based experiments

⚠️ Notes

• All synthetic samples in v1.0 were generated with OmniVoice, so results may reflect generator-specific artifacts.
• Emotion distribution is imbalanced, particularly Fear with only 17 pairs.
• Emotion labels correspond to the target emotional category; perceptual emotion preservation has not been independently validated.

FaEmoSpoof v1.0
5,696 Persian emotional speech samples · 2,848 matched Real/Fake pairs