Datasets:

lab260
/

LRLspoof

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
size_categories:
  - 1M<n<10M
task_categories:
  - audio-classification
pretty_name: LRL spoof
tags:
  - antispoofing

LRLspoof: Low-Resource Language Spoofing Corpus

LRLspoof is a large-scale multilingual synthetic-speech corpus designed for cross-lingual spoof detection.

The dataset was introduced in the paper When Spoof Detectors Travel: Evaluation Across 66 Languages in the Low-Resource Language Spoofing Corpus.

Dataset Summary

Total Audio: 2,732 hours
Number of Languages: 66 (including 45 low-resource languages)
Generation: Audio was generated using 24 different open-source Text-to-Speech (TTS) systems.
Purpose: Benchmarking countermeasures for synthetic speech detection (spoofing) across diverse linguistic domains.

Key Features

Multilingual Scope: Covers a wide variety of languages to evaluate how language-specific features act as a source of domain shift in spoof detection.
Scale: One of the largest available corpora for synthetic speech detection, facilitating robust training and evaluation.
Low-Resource Focus: Explicitly includes 45 languages defined as low-resource, addressing a gap in current AI safety and security research.