Real and cloned voice speech data for deepfake detection
Real voices next to their clones: consenting speakers recorded live in English, Hindi, Arabic and Portuguese, then imitated by text-to-speech, voice-conversion and voice-cloning systems, and played back over real phones, networks and speakers. Every clip is labelled real or fake, with its attack and channel, for detectors that have to work on calls, not only in the lab.
Specification
- Languages
- English (US, UK and Indian), Hindi, Arabic and Brazilian Portuguese
- Speakers
- Consenting speakers who agree to their voices being cloned for this purpose, balanced across gender and age groups
- Real and fake
- Live read and spontaneous speech, and text-to-speech, voice-conversion, voice-cloning and replay attacks of the same voices
- Channels
- Mobile and landline calls, VoIP and messaging apps, speaker-to-phone replay, compression and noise
- Annotation
- Real or fake, attack type, the kind of system used, channel, codec and noise conditions
- Quality check
- Reviewers confirm every real clip and every attack label
Example of the data
What you see, and the file you get.
| clip 0118 | real | mobile call |
|---|---|---|
| clip 0119 | fake · voice clone | same speaker, VoIP |
| clip 0120 | fake · replay | speaker → phone |
| audio_url | speaker | language | label | attack | channel | codec |
|---|---|---|---|---|---|---|
| …/audio/clip_0118.wav | S041 | en-IN | real | mobile_call | amr_nb | |
| …/audio/clip_0119.wav | S041 | en-IN | fake | voice_clone | voip | opus |
| …/audio/clip_0120.wav | S041 | en-IN | fake | replay | speaker_to_phone | amr_nb |
| …/audio/clip_0231.wav | S077 | ar-SA | fake | tts | landline | g711 |
| …/audio/clip_0232.wav | S077 | ar-SA | real | messaging_app | opus |
Sample rows. Also as JSON or Excel; audio as WAV or MP3.
Contact us to get the dataset
Tell us what you need. We walk you through the data, its formats and how it is checked on a short call.
What it is used for
Voice-biometric security
Stop cloned voices from passing voice log-in at banks, telcos and help desks.
Call fraud detection
Flag synthetic callers impersonating customers or executives.
Deepfake detection across languages
Train detectors that hold up in Hindi, Arabic and Portuguese, 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.
Why replay over real phones?
Because attacks arrive through phone lines, apps and codecs that change the audio, and detectors trained only on clean files miss them.
Can I see samples first?
Yes. Request samples with the form on this page and we will send them with the full specification.
Tell us what your model needs.
Request samples and the full specification. We reply within a couple of business days.
Request samples