SEA-Spoof

SEA-Spoof is a multilingual speech anti-spoofing dataset for audio deepfake detection, with paired bonafide and spoof speech and transcript-aligned metadata. The dataset is designed for controlled language-level evaluation of spoofed speech detection, especially for South-East Asian language settings.

This Hugging Face release is a cleaned dataset package with embedded audio bytes, transcript text, labels, language tags, and text provenance. Server-local absolute paths are not stored in the dataset records.

Abstract

The rapid growth of the digital economy in South-East Asia (SEA) has amplified the risks of audio deepfakes, yet existing datasets provide limited coverage of SEA languages, hindering robust detection. We present SEA-Spoof, the first large-scale audio deepfake detection dataset dedicated to six SEA languages: Tamil, Hindi, Thai, Indonesian, Malay, and Vietnamese. SEA-Spoof contains over 700 hours of paired real and spoof speech generated by diverse state-of-the-art open-source and closed-source systems. Its balanced, transcript aligned design enables controlled language and system level evaluation. Benchmarking reveals severe cross-lingual degradation of models trained on high resource languages, while fine-tuning on SEA-Spoof restores performance across languages and synthesis sources. SEA-Spoof establishes a foundation for robust, cross-lingual and region-aware deepfake detection in SEA.

Dataset Size

The cleaned HF package contains 553,746 utterances in 38 Parquet shards.

Split	Utterances
train	439,362
validation	57,158
evaluation	57,226
total	553,746

Label distribution:

Label	Utterances	Percent
spoof	294,964	53.27%
bonafide	258,782	46.73%

Language coverage:

Language code	Language	Utterances	Percent
en	English	130,776	23.62%
hi	Hindi	71,171	12.85%
id	Indonesian	93,542	16.89%
ms	Malay	42,382	7.65%
ta	Tamil	49,640	8.96%
th	Thai	85,215	15.39%
vi	Vietnamese	81,020	14.63%

High-level category distribution:

Category	Utterances	Percent
bonafide	258,782	46.73%
offline_spoof	215,539	38.92%
online_spoof	79,425	14.34%

All audio records are stored as FLAC bytes with sampling_rate=16000.

Utterance Fields

Each row corresponds to one utterance.

Field	Description
row_id	Globally unique row identifier for this HF release.
utterance_id	Original SEA-Bench utterance identifier. This may not be globally unique across all source subsets, so use row_id when a unique key is required.
audio	Embedded audio object with FLAC bytes and a clean relative logical path.
text	Transcript text associated with the audio.
language	Language code: en, hi, id, ms, ta, th, or vi.
label	bonafide or spoof.
spoof_type	Spoof subtype or bonafide.
category	Higher-level data category, such as bonafide, offline_spoof, or online spoof subsets.
split	Dataset split: train, validation, or evaluation.
text_source	Detailed provenance for the transcript, such as a local manifest/metadata source or WhisperX backfill.
mapping_source	Transcript source group: local or whisperx_backfill.
text_granularity	Indicates whether the transcript came from utterance-level or source/video-level metadata when applicable.
is_text_exact	Whether the transcript is expected to be an exact utterance-level match.
sampling_rate	Audio sampling rate, fixed at 16,000 Hz in this package.
audio_was_resampled	Whether the audio was resampled during packaging.

Transcript Provenance

Most transcripts come from local SEA-Spoof/SEA-Bench metadata, generation manifests, or transcript bundles. A smaller part was backfilled using WhisperX for audio without usable local text.

mapping_source	Utterances	Percent	Meaning
local	498,284	89.98%	Transcript found in local metadata, transcript files, or generation manifests.
whisperx_backfill	55,462	10.02%	Transcript generated with WhisperX for rows that did not have usable local transcript text.

To inspect where a transcript came from, read the mapping_source, text_source, text_granularity, and is_text_exact fields. For example, mapping_source="local" means the text was recovered from existing local metadata/transcript resources. mapping_source="whisperx_backfill" means the text was produced during the ASR backfill step.

Usage Examples

Install:

pip install datasets soundfile

Load the dataset:

from datasets import load_dataset

ds = load_dataset("Jack-ppkdczgx/SEA-Spoof")
print(ds)
print(ds["train"][0])

Stream examples without downloading all shards first:

from datasets import load_dataset

train = load_dataset("Jack-ppkdczgx/SEA-Spoof", split="train", streaming=True)
example = next(iter(train))

print(example["row_id"])
print(example["language"], example["label"])
print(example["text"])
print(example["mapping_source"], example["text_source"])

Select one language, for example Vietnamese:

from datasets import load_dataset

train = load_dataset("Jack-ppkdczgx/SEA-Spoof", split="train", streaming=True)
vi_train = train.filter(lambda x: x["language"] == "vi")

for example in vi_train.take(3):
    print(example["row_id"], example["label"], example["text"][:120])

Select Malay spoof examples:

from datasets import load_dataset

train = load_dataset("Jack-ppkdczgx/SEA-Spoof", split="train", streaming=True)
ms_spoof = train.filter(lambda x: x["language"] == "ms" and x["label"] == "spoof")

for example in ms_spoof.take(3):
    print(example["row_id"], example["category"], example["spoof_type"])

Compare local transcripts and WhisperX-backfilled transcripts:

from datasets import load_dataset

train = load_dataset("Jack-ppkdczgx/SEA-Spoof", split="train", streaming=True)

local_text = train.filter(lambda x: x["mapping_source"] == "local")
asr_text = train.filter(lambda x: x["mapping_source"] == "whisperx_backfill")

print(next(iter(local_text))["text_source"])
print(next(iter(asr_text))["text_source"])

Read audio:

from datasets import load_dataset
import io
import soundfile as sf

ds = load_dataset("Jack-ppkdczgx/SEA-Spoof", split="validation", streaming=True)
example = next(iter(ds))

# The audio is embedded as FLAC bytes in each row.
audio_bytes = example["audio"]["bytes"]
array, sampling_rate = sf.read(io.BytesIO(audio_bytes))
print(array.shape, sampling_rate)
print(example["text"])

If your environment decodes the column as a Hugging Face Audio feature, this form is also available:

from datasets import load_dataset

ds = load_dataset("Jack-ppkdczgx/SEA-Spoof", split="validation", streaming=True)
example = next(iter(ds))

array = example["audio"]["array"]
sampling_rate = example["audio"]["sampling_rate"]
print(array.shape, sampling_rate)

Intended Uses

SEA-Spoof is intended for academic research on:

• multilingual audio deepfake detection
• spoofed versus bonafide speech classification
• cross-lingual and language-specific robustness evaluation
• transcript-aware speech anti-spoofing analysis

The dataset is not intended for commercial use, speaker impersonation, voice cloning, surveillance, biometric deployment, or any harmful audio generation or misuse.

Access And License

This dataset is released for non-commercial academic research only.

Use is restricted to academic institutions and approved research users. Commercial use is not permitted, and this dataset may not be used by commercial companies or for commercial products, services, model training, evaluation, or deployment.

Access requires author approval. To request access, please email:

• wu_jinyang@a-star.edu.sg
• imcc.sg@gmail.com

Please include your name, affiliation, intended research use, and whether the use is academic and non-commercial.

Citation

If SEA-Spoof is interesting or useful for your research, please cite our paper. We also welcome discussion and academic collaboration.

@article{wu2025sea,
  title={SEA-Spoof: Bridging The Gap in Multilingual Audio Deepfake Detection for South-East Asian},
  author={Wu, Jinyang and Hou, Nana and Pan, Zihan and Zhang, Qiquan and Bhupendra, Sailor Hardik and Mondal, Soumik},
  journal={arXiv preprint arXiv:2509.19865},
  year={2025}
}

Contact

For access requests, questions, or collaboration discussions, please contact:

• wu_jinyang@a-star.edu.sg
• imcc.sg@gmail.com