ANC-Spoof: Audio Neural Codec Spoof Dataset

Overview

ANC-Spoof is a large-scale dataset for studying the robustness of audio deepfake detection (ADD) systems against distortions introduced by neural audio codecs. It pairs original (uncompressed) audio from three established ADD benchmarks with codec-resynthesized versions of the same utterances, produced by eight different neural codecs (including the uncompressed Original version).

The dataset is built from:

• ASVspoof 2019 LA — used for training, validation (dev), and evaluation
• Fake-or-Real (FoR) — used for cross-dataset evaluation only
• In-the-Wild — used for cross-dataset evaluation only

Dataset Structure

The dataset is distributed as three configs, each independently downloadable:

Config	Splits	Rows
ASVspoof2019	train, dev, eval	971,688
FoR	data	37,072
InTheWild	eval	254,232

Only ASVspoof2019 has train/dev splits — FoR and InTheWild are provided purely as out-of-domain generalization test sets, following standard practice in ADD research (train on ASVspoof19, evaluate cross-dataset on FoR/In-the-Wild).

Each row represents one (utterance, codec) pair:

Column	Type	Description
audio	Audio	The waveform, embedded (array + sampling rate)
label	ClassLabel	0 = bonafide (real), 1 = spoof (fake)
codec	string	Which processing was applied: Original, BigCodec, DAC, HiggsAudioV2, Mimi, SNAC, SpeechTokenizer, WavTokenizer
subset	string	Source benchmark: ASVspoof2019, FoR, InTheWild
split	string	train / dev / eval
speaker	string or null	Speaker ID, where available
utt_id	string	Original utterance identifier

Usage

from datasets import load_dataset

# Load only what you need — each call fetches just that config/split
asv_train = load_dataset("abdulahh35/ANC-Spoof", "ASVspoof2019", split="train")
asv_dev   = load_dataset("abdulahh35/ANC-Spoof", "ASVspoof2019", split="dev")
asv_eval  = load_dataset("abdulahh35/ANC-Spoof", "ASVspoof2019", split="eval")

for_eval  = load_dataset("abdulahh35/ANC-Spoof", "FoR", split="data")
wild_eval = load_dataset("abdulahh35/ANC-Spoof", "InTheWild", split="eval")

# Filter to a single codec, e.g. only BigCodec-compressed utterances
bigcodec_only = asv_eval.filter(lambda x: x["codec"] == "BigCodec")

row = asv_dev[0]
waveform, sr = row["audio"]["array"], row["audio"]["sampling_rate"]
label_str = asv_dev.features["label"].int2str(row["label"])  # "bonafide" or "spoof"

Label Encoding

ds.features["label"].names  # ['bonafide', 'spoof']
# 0 -> bonafide (real)
# 1 -> spoof (fake)

Source Datasets

Please also cite the original benchmark datasets this data was derived from:

• ASVspoof 2019 (Wang et al., 2019)
• In-the-Wild (Müller et al., 2022)
• Fake-or-Real / FoR (Reimao & Tzerpos, 2019)

License

Citation