ARFAKE: A Multi-Dialect Benchmark and Baselines for Arabic Spoof-Speech Detection

Overview

ARFAKE is the first multi-dialect Arabic spoof-speech benchmark, designed to evaluate and advance anti-spoofing systems for Arabic audio. With the rapid progress of generative text-to-speech (TTS) and voice-cloning models, distinguishing between real and synthetic speech has become increasingly challenging, especially for Arabic and its diverse dialects — a language family that has been underrepresented in previous deepfake detection .

This repository provides:

• The ARFAKE dataset, built on top of the Casablanca speech corpus (8 dialects, ~6 hours each).
• Spoofed versions generated using state-of-the-art TTS systems:
  • XTTS-v2
  • FishSpeech
  • ArTST
  • VITS
• Baselines and evaluation pipeline for detecting spoofed speech using both traditional ML and modern embedding-based models.

---

Key Features

• 📀 Multi-dialect coverage: Eight Arabic dialects, balanced across bonafide and spoofed samples.
• 🎙️ Spoofed data generation: Using large-scale multilingual and Arabic-specific TTS models.
• 🧪 Detection baselines:
  • MFCC + classical ML classifiers (SVM, Random Forest, etc.)
  • Embedding-based models using HuBERT, Whisper, and Wav2Vec 2.0
  • RawNet2, the ASVspoof benchmark system
• 🔍 Evaluation metrics:
  • Equal Error Rate (EER)
  • Accuracy
  • Mean Opinion Score (MOS) (via human ratings)
  • Word Error Rate (WER) (via Whisper-Large ASR)

---

Dataset

• Source corpus: [Casablanca dataset (2024)]
• Size: 54,413 utterances (~23k test samples, ~31k train samples)
• Composition:
  • Bonafide (genuine) speech
  • Spoofed speech from FishSpeech, XTTS-v2, and ArTST
• Dialectal coverage: DZ, EG, JO, MA, MR, PS, AE, YE (ISO 3166-1 alpha-2 codes)
• Distribution: (see Figure 1 in paper).

---

Baselines & Results

• Embedding-based models outperform traditional MFCC-based ML classifiers.
• Whisper-large achieved the best detection performance (EER 6.92% on FishSpeech-generated data).
• FishSpeech produced the most challenging spoofed samples, with the highest MOS (3.72/5) and lowest WER, making it harder to detect than XTTS-v2, ArTST, or VITS.
• Classifiers trained on the combined dataset generalized well even to unseen TTS models like VITS.

Summary of Findings:

• FishSpeech is the most realistic and difficult TTS system for Arabic spoofing.
• Combining spoofed data from multiple TTS models improves generalizability of detectors.
• Whisper-based detectors outperform MFCC-based baselines by a wide margin.

---

Usage

1. Dataset Access
   We uploaded the dataset, you can find use merge_training_set to train your model and merge_test_set (in-domain) ,Vits-spoofed (out-domain).

2. Training Baseline Models

   • Classical ML: Train SVM, Random Forest, etc. on MFCC features.
   • Embedding-based: Use pre-trained HuBERT / Whisper / Wav2Vec encoders with classifier heads.
   • Benchmark comparison with RawNet2.

3. Evaluation

   • Run detection and report EER, Accuracy, MOS, and WER.
   • Use Whisper-Large for ASR-based evaluation.

---

Citation

@misc{maged2025arfakemultidialectbenchmarkbaselines,
      title={ArFake: A Multi-Dialect Benchmark and Baselines for Arabic Spoof-Speech Detection}, 
      author={Mohamed Maged and Alhassan Ehab and Ali Mekky and Besher Hassan and Shady Shehata},
      year={2025},
      eprint={2509.22808},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.22808}, 
}