IlyasDaoud/real-or-clone-data

The Real or Clone? dataset

English and French speech, real vs AI-cloned, every clip turned into a WhatsApp-style voice note. It trained and tested the detector published at IlyasDaoud/real-or-clone-xlsr.

Why there is no audio download. The dataset is made of clips from four public datasets (two of them gated or non-commercial) plus our own clones, and it lived on the hackathon GPU box, which has since been deleted. What is published here is the exact recipe: every choice is seeded, so the scripts below select the same source clips, the same speaker and generator splits and the same voice-note effects again. Only our Chatterbox clones come out slightly different, because voice generation is random by nature.

What's inside (run C)

	Real	Clone	Total
Train	24,670	26,344	51,014
Test (held out)	4,120	6,002	10,122

Languages: English and French only. Labels: 0 = real, 1 = clone.

Source	Label	What we took	Licence
FLEURS	real	en_us + fr_fr, all splits (read speech)	CC-BY-4.0
VoxPopuli	real	4,000 clips per language (parliament speech, many speakers and mics)	CC0
In-the-Wild	real + clone	English celebrity and politician speech, real and deepfaked	CC-BY-SA-4.0
MLAAD	clone	150 French and 120 English files per generator (180+ TTS systems)	CC-BY-NC-4.0, gated
Our clones (generate/clone.py)	clone	1,619 Chatterbox Multilingual clones of FLEURS speakers, scam-style texts	same as FLEURS + non-commercial use

Real voices come from several sources on purpose: with FLEURS alone, run A learned "sounds like FLEURS = real" (see the README's Results).

Splits: nothing in the test set was seen in training

Source	How it is split
FLEURS	its own train + dev → train, its test → test
VoxPopuli	by speaker: 20% of speakers test-only
In-the-Wild	by speaker: 30% of speakers test-only
MLAAD	by generator: every commercial API (ElevenLabs, OpenAI, Gemini, Cartesia, MiniMax, Hume, Resemble…) plus ~20% of the others are test-only
Our clones	by reference speaker: 20% test-only

The voice-note simulation

Every clip, real and fake, goes through the same recipe, so audio quality can't reveal the label (augment/voicenote.py):

16 kHz mono → trim edge silence → random 8 s crop → random loudness → one condition, picked from a hash of the file path (never from the label): clean, Opus 16 kbps, phone (300–3400 Hz band, 8 kHz, Opus 12 kbps), noise + Opus or reverb + Opus.

Layout

$DATA_DIR/                     (default ~/roc_data)
  fleurs/<lang>/<split>/*.wav      itw/release_in_the_wild/*.wav + meta.csv
  mlaad/fake/<lang>/<generator>/   voxpopuli/<lang>/*.wav
  clones/*.wav                     manifest.csv      (every source clip: path, label, source, generator, lang, split)
  vn/<hash>.wav                    manifest_vn.csv   (the voice-note versions the model trains and tests on)

Rebuild it

On a Linux box with an NVIDIA GPU (the clones need one; we used an L40S on NVIDIA Brev, a few hours in total) and about 60 GB of disk. MLAAD is gated: accept its terms on Hugging Face first and put a read token in .env.

git clone https://github.com/ilyasdaoudrma/real-or-clone && cd real-or-clone
sudo apt-get install -y ffmpeg
python3 -m venv ~/venv-main && source ~/venv-main/bin/activate && pip install -r requirements.txt
cp .env.example .env                       # HF_TOKEN=<read token with MLAAD access>

python data/download.py                    # FLEURS, In-the-Wild (8.2 GB zip), MLAAD FR/EN, VoxPopuli
python3 -m venv ~/venv-tts && ~/venv-tts/bin/pip install chatterbox-tts "setuptools<81"
for i in 0 1 2 3 4 5; do ~/venv-tts/bin/python generate/clone.py --n 500 --shard $i --num-shards 6 & done; wait
python data/build_manifest.py              # -> manifest.csv with the splits above
python augment/voicenote.py                # -> vn/*.wav + manifest_vn.csv

Smoke test on a laptop, a few files per source: python data/download.py --tiny.

The selection is identical as long as the source datasets are unchanged; MLAAD keeps adding generators, so a rebuild done much later can pick up new ones.

Use and ethics

Research and non-commercial use only (MLAAD is CC-BY-NC-4.0, In-the-Wild is share-alike). Our clones are synthetic copies of FLEURS speakers' voices, made only to train a detector; don't use them to impersonate anyone.