README.md

metadata

language:
  - ab
  - hy
  - as
  - av
  - az
  - ba
  - be
  - pt
  - bg
  - ca
  - ce
  - cv
  - hr
  - cs
  - da
  - nl
  - en
  - fa
  - fi
  - fr
  - ka
  - gu
  - hi
  - hu
  - is
  - id
  - it
  - ja
  - kk
  - ko
  - ky
  - lv
  - lt
  - lb
  - ms
  - ml
  - mr
  - ne
  - or
  - os
  - pl
  - ru
  - sl
  - es
  - sw
  - te
  - tk
  - uz
  - cy
  - aus
  - brx
  - arr
  - crh
  - myv
  - xal
  - krc
  - kjh
  - lez
  - mni
  - mhr
  - mdf
  - nog
  - raj
  - tt
  - tyv
  - sah
license: mit
task_categories:
  - audio-classification
tags:
  - antispoofing
  - text-to-speech
  - anti-spoofing
  - audio-deepfake-detection
  - speech
  - benchmark
  - arena-ready
pretty_name: LRL spoof
size_categories:
  - 1M<n<10M
configs:
  - config_name: default
    data_files:
      - split: test
        path: lrl_spoof.tar.gz.part_*
viewer: false
arxiv:
  - '2603.02364'

LRLspoof

Official dataset for our INTERSPEECH 2026 paper "When Spoof Detectors Travel: Evaluation Across 66 Languages in the Low-Resource Language Spoofing Corpus" (arXiv:2603.02364). This repository hosts the benchmark, evaluation protocol, and baseline submissions released with the paper. If you use LRLspoof, please cite it.

Multilingual low-resource-language anti-spoofing benchmark: 1,304,455 TTS (spoof) utterances across 66 languages and many TTS systems. Spoof-only (no bonafide).

The audio ships as the multi-part tarball lrl_spoof.tar.gz.part_*; download + extract it into a <language>/<tts_model>/<file>.wav tree. load_dataset is not the access path here (the dataset is the tarball). data/labels.parquet (utterance_id + label, all spoof) exists only for arena re-verification.

Arena scoring

Scored on 1-SRR (srr_complement, lower is better) at a fixed threshold t* calibrated on DeepVoice. A submission must include a calibration block with the DeepVoice operating-point threshold.

utterance_id convention

utterance_id = the audio file path relative to the dataset root, e.g. english/fastpitch/line_59.wav. A submitter's scores.txt keys by this id; emit <utterance_id> <score> with higher score = more bonafide.

Citation

If you use LRLspoof, please cite our INTERSPEECH 2026 paper:

@inproceedings{borodin2026lrlspoof,
  title     = {When Spoof Detectors Travel: Evaluation Across 66 Languages
               in the Low-Resource Language Spoofing Corpus},
  author    = {Borodin, Kirill and Kudryavtsev, Vasiliy and Maslov, Maxim
               and Gorodnichev, Mikhail and Mkrtchian, Grach},
  booktitle = {Proc. INTERSPEECH 2026},
  year      = {2026},
  note      = {arXiv:2603.02364},
  url       = {https://arxiv.org/abs/2603.02364}
}

Contact

• Email: kborodin.research@gmail.com
• Telegram: @korallll_ai