Arabic voice anti-spoofing data: real and cloned speech
Real Arabic voices next to their clones: the same consenting speakers recorded live and then imitated by text-to-speech and voice-conversion systems, over phone lines and replayed through speakers. Labelled real or fake, by attack type, for detectors that protect banks and call centres.
Specification
- Speakers
- Consenting Saudi, Jordanian and Emirati speakers, who agree to their voices being cloned for this purpose
- Real speech
- Read and spontaneous speech, over phone lines and wideband microphones
- Fake speech
- Text-to-speech, voice conversion and replay attacks of the same voices
- Annotation
- Real or fake, attack type, the system used, channel and noise conditions
- Quality check
- Reviewers confirm every real clip and every attack label
Example of the data
| clip 0041 | real | phone line |
|---|---|---|
| clip 0042 | fake · voice conversion | same speaker |
| clip 0043 | fake · replay | speaker → phone |
What it is used for
Voice-biometric security
Stop cloned voices from passing voice log-in at banks and telcos.
Call-centre fraud detection
Flag synthetic callers impersonating customers.
Deepfake detection
Train detectors that hold up on Arabic dialects, not only English.
Red-teaming voice systems
Measure how your speaker verification fails, before attackers do.
Questions
Do the speakers agree to being cloned?
Yes. Every speaker gives informed consent for their recordings to be cloned for detection research and products.
Which attacks are included?
Text-to-speech and voice-conversion clones, and replay through speakers and phone lines. Your own attack systems can be added.
Can I see samples first?
Yes. Request samples with the form on this page and we will send them with the full specification.
Related datasets
Tell us what your model needs.
Request samples and the full specification. We reply within a couple of business days.
Request samples