SpeechAntiSpoofingBenchmarks/J-SPAW_LA

J-SPAW (LA track, eval)

⚠️ NON-COMMERCIAL USE ONLY

The upstream J-SPAW dataset is released "For non-commercial use only" (see the J-SPAW repository). This packaging inherits that restriction: do not use it for any commercial purpose. It is provided solely for non-commercial academic research and benchmarking. The upstream terms are sparse and do not spell out redistribution; contact the original authors for any use beyond non-commercial research.

Benchmark-ready packaging of the LA (logical access) track of J-SPAW — the Japanese Speaker-verification and sPoofing-attacks recorded in-the-Wild dataset (Shiota et al., Interspeech 2025), for speech anti-spoofing / deepfake voice detection.

Overview

J-SPAW is a Japanese, in-the-wild corpus built to evaluate speaker verification and spoofing detection jointly, with a focus on physical-access (replay) scenarios. This repository packages only the dataset's LA track, which is the subset shipped as audio in the ver1 release:

• bonafide — genuine human speech (iPhone 8 microphone M2), recorded across four acoustic environments (E1 quiet, E2 air-conditioner, E3 music, E4 outside).
• spoof — LA-track attacks L1 / L2.

Binary classification: bonafide (genuine) vs. spoof. The full J-SPAW PA (physical-access / replay) track is not part of this ver1 audio release and is therefore not packaged here.

3 source clips shorter than 1.0 s were dropped (the Arena requires ≥ 1.0 s).

License & redistribution

Non-commercial use only — see the banner above and LICENSE.txt. Labels and the evaluation protocol are unmodified; audio is the original 16 kHz mono PCM WAV, embedded bit-exactly (no re-encode — a full decode probe of all 2,400 source clips passed cleanly).

Schema

Column	Type	Description
path	string	LA/<uid>.wav (spoof) or ASV/<uid>.wav (bonafide), unique
audio	Audio(16000)	16 kHz mono WAV
label	ClassLabel	"bonafide" (0) / "spoof" (1)
notes	string	JSON: utterance_id, speaker_id, track, attack, record_env, mic, subset

notes example:

{"utterance_id": "F001_R1_E2_L1_BT", "speaker_id": "F001", "track": "LA", "attack": "L1", "record_env": "E2", "mic": null, "subset": "eval"}

Quick Start

from datasets import load_dataset

ds = load_dataset("SpeechAntiSpoofingBenchmarks/J-SPAW_LA", split="test")
print(ds[0])

Stats

Stat	Value
Total trials	2,397
Bonafide	800
Spoof	1,597
Attacks	L1 (797), L2 (800)
Total audio	~1.49 h
Language	Japanese

Source provenance

• Paper: S. Shiota, S. Horie, K. Kanno, S. Takamichi, "J-SPAW: Japanese speaker verification and spoofing attacks recorded in-the-wild dataset", Interspeech 2025. DOI 10.21437/Interspeech.2025-352.
• Dataset: https://github.com/takamichi-lab/j-spaw
• Protocol: metadata_LA.txt (ASVspoof-2021-style LA-track protocol).

Evaluation

For evaluation instructions and submission format, see submissions/README.md.

Citation

@inproceedings{shiota2025jspaw,
  title     = {{J-SPAW: Japanese speaker verification and spoofing attacks recorded in-the-wild dataset}},
  author    = {Shiota, Sayaka and Horie, Suzuka and Kanno, Kouta and Takamichi, Shinnosuke},
  year      = {2025},
  booktitle = {Proc. Interspeech 2025},
  pages     = {3913--3917},
  doi       = {10.21437/Interspeech.2025-352},
}

Maintainer

Contact: k.n.borodin@mtuci.ru