NAUSS-Spoof50: A Multi-Task Arabic Speech Benchmark for Audio Spoofing Detection and Forgery Localization

Metadata and Documentation Release

Version: 1.0
License: Creative Commons Attribution Non-Commercial 4.0 (CC BY-NC 4.0)
Access: Metadata freely accessible. Audio files available upon request.

---

Overview

NAUSS-Spoof50 is a curated multi-task Arabic speech benchmark designed to support both fully synthetic spoofing detection and partially manipulated forgery localization within a unified evaluation framework. The dataset comprises 76,730 utterances from 50 native Arabic speakers with balanced gender distribution, recorded across three consumer-grade devices and two acoustic environments under a strict speaker-disjoint protocol.

This release provides metadata files and dataset documentation to support transparency and reproducibility. Raw audio files are not included in this public release due to participant privacy constraints and institutional security considerations. Audio access is available through the gated access procedure described below.

---

Dataset Summary

Property	Value
Total utterances	76,730
Native speakers	50 (25 male, 25 female)
Recording devices	3 (mobile, laptop, microphone)
Acoustic environments	2 (clean, normal)
Spoofing paradigms	6 (HiFiGAN, BigVGAN, Encodec, MMS-Arabic, XTTS v2, ElevenLabs)
Bonafide utterances	16,200
Spoofed utterances	60,530
Forgery localization files	16,200
Annotation resolution	10 ms frame-level masks
Sampling rate	16 kHz
Utterance duration	5 seconds (fixed)
Language	Arabic
Split protocol	Speaker-disjoint (train/dev/eval)

---

Dataset Splits

Split	Speakers	Bonafide	Spoofed	Total
Train	35	10,035	44,267	54,302
Dev	7	2,915	10,608	13,523
Eval	8	3,250	5,655	8,905
Total	50	16,200	60,530	76,730

---

Speaker Organization

Speakers are assigned globally unique IDs (SPK001–SPK050) using a gender-grouped scheme:

Group	Speaker IDs	Count
Batch A Female	SPK001–SPK015	15
Batch A Male	SPK016–SPK030	15
Batch B Female	SPK031–SPK040	10
Batch B Male	SPK041–SPK050	10
Total	SPK001–SPK050	50

---

Spoofing Paradigms

Generator	Paradigm	Total
HiFiGAN	Neural vocoder	10,942
BigVGAN	Neural vocoder	10,942
Encodec	Neural codec	10,957
MMS-Arabic	Text-to-speech	2,775
XTTS v2	Voice cloning	6,914
ElevenLabs	Voice cloning	18,000
Total		60,530

---

Forgery Localization Subset

Transform	Total Segments
Copy-paste	8,709
Pitch shift	7,441
Time warp	7,286
Local shift	7,267
Partial delete	5,614
Total	16,200 files / 43,800+ segments

---

Files in This Release

File	Description
bonafide_metadata.csv	Metadata for all 16,200 bonafide recordings
spoof_metadata.csv	Metadata for all 60,530 spoofed recordings
forgery_metadata.csv	Metadata for all 16,200 forgery localization files
README.md	This file
DATA_DICTIONARY.md	Column descriptions for all CSV files

---

File Naming Conventions

Bonafide:

NAUSS50_BF_{speaker_id}_{gender}_{environment}_{device}_{utterance_index}
Example: NAUSS50_BF_SPK010_female_clean_laptop_001

Spoofed:

NAUSS50_SP_{spoof_type}_{split}_{index}
Example: NAUSS50_SP_HiFiGAN_train_00001
         NAUSS50_SP_ElevenLabs_eval_00042

Forgery:

NAUSS50_FG_{split}_{index}
Example: NAUSS50_FG_train_00001
         NAUSS50_FG_eval_00042

---

Evaluation Protocols

Protocol 1 — Mixed Attack Detection: All spoofed samples combined regardless of generation method. Supports aggregate EER and minDCF evaluation.

Protocol 2 — Generator-Wise Analysis: Spoof detection evaluated separately per spoofing method. Supports per-generator EER comparison using the spoof_type field in spoof_metadata.csv.

Both protocols use the speaker-disjoint split. No speaker appears in more than one split.

---

Ethics and Privacy

This dataset was collected under approval from the Ethics Committee of Naif Arab University for Security Sciences (NAUSS) in 2024. Informed consent was obtained from all participants prior to recording. Due to participant privacy constraints and institutional security considerations, raw audio files are not publicly distributed.

This access model follows established practice for sensitive human speech and biometric datasets, including Bridge2AI Voice, TAME Pain, CHiME-9 ECHI, and the Speech Accessibility Project.

---

Requesting Full Dataset Access

Access to the complete NAUSS-Spoof50 audio dataset is available for non-commercial academic research through the gated access form on this page. Requests require:

1. Name and institutional affiliation
2. Brief description of intended research use
3. Agreement to the Data Use Agreement (DUA)

Requests are reviewed within 30 days. The DUA prohibits redistribution, commercial use, and any attempt to identify participants from voice recordings.

---

Citation

If you use this metadata release or the NAUSS-Spoof50 benchmark in your research, please cite:

Moallim, M.; Alhaj, T.A.; Elhaj, F.A.; Darwish, T.
NAUSS-Spoof50: A Multi-Task Arabic Speech Benchmark for
Audio Spoofing Detection and Forgery Localization.
Signals, 2026. (Under Review)

---

Contact

Corresponding author: Dr. Taqwa Ahmed Alhaj
Email: talhaj@nauss.edu.sa
Institution: Center of Artificial Intelligence for Security, Naif Arab University for Security Sciences (NAUSS), Riyadh 14812, Saudi Arabia