mueller91/WASP

Paper: https://arxiv.org/pdf/2606.23335

WASP: Watermark-Shortcut Paired Speech Corpus

WASP is a paired speech corpus for studying the watermark shortcut in audio deepfake detection: the spurious "watermark => fake" feature a detector picks up when synthetic speech is watermarked by default and human speech is not. Every utterance is provided clean and under each watermark, all applied post hoc, so the clean and watermarked versions of an utterance differ only in the mark. Any change in a detector's score between the two members of a pair is therefore attributable to the watermark alone, with no confound from the generator.

This is the released corpus for the black-box study in the paper The Watermark Shortcut: How Provenance Marking Sabotages Audio Deepfake Detection.

Structure

{label}/{lang}/{source}/{model}/{watermark}/{utterance}.wav
e.g. spoof/en/mailabs/chatterbox/perth/ozma_of_oz_10_f000019.wav
     spoof/en/mailabs/chatterbox/without/ozma_of_oz_10_f000019.wav   # same base, no mark

• label: spoof (TTS output) or bonafide (genuine human speech).
• lang: en, de, fr, zh.
• source: mailabs (M-AILABS; en/de/fr) or aishell (AISHELL-3; zh). For bonafide, the model level is the literal bonafide.
• model: the TTS system (spoof only); see below.
• watermark: without (clean base), perth, wavmark, audioseal, silentcipher; bonafide additionally carries all-WMs (all marks stacked).

Pairing. A watermarked clip and its clean counterpart share the same {label}/{lang}/{source}/{model}/{utterance} and differ only in the {watermark} directory; the clean base lives under without/.

Synthesizers (spoof)

Model	Default watermark
Chatterbox	PerTh
Chatterbox-Turbo	PerTh
DramaBox	PerTh
Kyutai	none
Orpheus	none (SilentCipher opt-in)
Sesame CSM	SilentCipher

The built-in watermark of each model is disabled during generation to obtain a clean base; all marks in the corpus are then applied post hoc. English carries all six systems; German, French, and Mandarin are Chatterbox-led.

Watermark schemes

PerTh (Resemble AI), WavMark, AudioSeal (Meta), and SilentCipher (Sony), plus an all-WMs stack for the bona-fide split. All operate on a finished waveform.

Size

About 7,000 clips derived from roughly 1,400 base utterances (~1,000 synthetic, 400 genuine), across four languages.

Loading

# Build a Hugging Face datasets object with metadata parsed from the path
from huggingface_hub import snapshot_download
from datasets import Dataset, Audio
import glob, os

local = snapshot_download("mueller91/WASP", repo_type="dataset")
rows = []
for wav in glob.glob(os.path.join(local, "**", "*.wav"), recursive=True):
    label, lang, source, model, watermark = os.path.relpath(wav, local).split(os.sep)[:5]
    rows.append({"audio": wav, "label": label, "lang": lang, "source": source,
                 "model": model, "watermark": watermark,
                 "utterance": os.path.basename(wav)})
ds = Dataset.from_list(rows).cast_column("audio", Audio())
print(ds[0])

Download

# Option A - Hugging Face CLI
huggingface-cli download mueller91/WASP --repo-type dataset --local-dir ./WASP

# Option B - huggingface_hub (Python)
from huggingface_hub import snapshot_download
snapshot_download("mueller91/WASP", repo_type="dataset", local_dir="./WASP")

# Option C - git clone (requires git-lfs)
git lfs install
git clone https://huggingface.co/datasets/mueller91/WASP

Intended use and license

Research on deepfake detection, watermark robustness, and shortcut learning. Source audio derives from M-AILABS and AISHELL-3; downstream use is subject to those corpora's terms in addition to the synthesizers' and watermark schemes' licenses. Use for building or evaluating systems intended to deceive or to falsely accuse is out of scope.

Citation

@article{mueller2026watermarkshortcut,
  title   = {The Watermark Shortcut: How Provenance Marking Sabotages Audio Deepfake Detection},
  author  = {M\"uller, Nicolas M. and Debus, Pascal},
  journal = {arXiv preprint arXiv:2606.23335},
  year    = {2026},
  note    = {Submitted to IEEE Signal Processing Letters}
}