ArFake-Dataset

ARFAKE: A Robust Framework for Multi-Dialect Arabic Speech Spoofing Detection Benchmark

ARFAKE is the first end-to-end benchmark for Arabic speech spoofing detection across multiple dialects. The framework systematically generates synthetic Arabic speech, evaluates its intelligibility and realism, constructs a large-scale spoofing dataset, trains robust detectors, and evaluates generalization across both unseen generators and unseen dialects.

---

Overview

ARFAKE addresses a major gap in Arabic speech security research by providing:

• 8 Arabic dialects
• 4 Text-to-Speech (TTS) systems
• 54K+ utterances
• Spoofed and bona fide speech
• Cross-generator robustness evaluation (LOGO)
• Cross-dialect robustness evaluation (LODO)

---

Abstract

Recent advances in generative text-to-speech systems have made synthetic speech increasingly realistic, creating new challenges for speech security and deepfake detection. While substantial progress has been made for English, Arabic remains underrepresented despite its rich dialectal diversity.

ARFAKE introduces the first comprehensive framework for generating and detecting spoofed Arabic speech across eight dialects. Using multiple state-of-the-art TTS systems, we construct a large-scale Arabic anti-spoofing benchmark and evaluate detector robustness under both generator and dialect shifts.

Our benchmark establishes a scalable foundation for future research in Arabic speech security, deepfake detection, and low-resource language anti-spoofing.

---

Key Contributions

1. Multi-Dialect Arabic Spoofing Dataset

Using the Casablanca multi-dialect Arabic corpus, we generate synthetic speech across:

• Algeria (DZ)
• Egypt (EG)
• Jordan (JO)
• Morocco (MA)
• Mauritania (MR)
• Palestine (PS)
• UAE (AE)
• Yemen (YE)

2. Diverse TTS Attack Generators

The benchmark includes synthetic speech generated using:

• XTTS-v2
• FishSpeech
• ArTST
• VITS

These systems represent different synthesis architectures and levels of realism.

3. Robust Anti-Spoofing Detectors

We evaluate:

SSL-Based Models

• Whisper-Large
• Whisper-Small
• HuBERT-Base
• wav2vec2

Traditional Baselines

• MFCC + SVM
• Logistic Regression
• Random Forest
• Extra Trees
• AdaBoost
• KNN
• Decision Trees

ASVspoof Baseline

• RawNet2

4. Novel Robustness Protocols

In-Domain Evaluation

Train and test on seen generators.

LOGO (Leave-One-Generator-Out)

Train on multiple generators and evaluate on a completely unseen TTS system.

LODO (Leave-One-Dialect-Out)

Train on seven dialects and evaluate on an unseen Arabic dialect.

---

Dataset Statistics

Split	Samples
Train + Validation	31,302
Test	23,111
Total	54,413

Dataset composition:

• Bona fide speech
• FishSpeech spoofed speech
• XTTS-v2 spoofed speech
• ArTST spoofed speech

VITS is reserved for unseen-generator evaluation under LOGO.

---

Main Results

Evaluation Protocol	Best Performance
In-Domain	96.86% Accuracy
LOGO (Unseen Generator)	97.94% Accuracy
LODO (Unseen Dialect)	Up to 93.51% Accuracy

Whisper-Large achieved the strongest overall performance across benchmark settings.

---

Citation

If you use ARFAKE in your research, please cite:

@misc{elsetohy2026arfakerobustframeworkmultidialect,
      title={ArFake: A Robust Framework for Multi-Dialect Arabic Speech Spoofing Detection Benchmark}, 
      author={Mohamed Elsetohy and Alhassan Ehab and Ali Mekky and Besher Hassan and Shady Shehata},
      year={2026},
      eprint={2509.22808},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.22808v2}, 
}

Please Cite also casablanca as it was used in the dataset building

@article{talafha2024casablanca,
   title={Casablanca: Data and Models for Multidialectal Arabic Speech Recognition},
   author={Talafha, Bashar and Kadaoui, Karima and Magdy, Samar Mohamed and Habiboullah, Mariem
           and Chafei, Chafei Mohamed and El-Shangiti, Ahmed Oumar and Zayed,
           Hiba and Alhamouri, Rahaf and Assi, Rwaa and Alraeesi, Aisha and others},
   journal={arXiv preprint arXiv:2410.04527},
   year={2024}
   
}

---

Acknowledgements

This work builds upon the Casablanca multi-dialect Arabic speech corpus and several open-source speech synthesis and self-supervised learning models.

---

Status

• Paper submitted
• Dataset release
• Training code release
• Pretrained models release

Stay tuned for updates.