ASVspoof 2019 (A large-scale public database of synthesized, converted and replayed speech)

Authors:
Xin Wang, Junichi Yamagishi, Massimiliano Todisco, Hector Delgado, Andreas Nautsch, Nicholas Evans, Md Sahidullah, Ville Vestman, Tomi Kinnunen, Kong Aik Lee, Lauri Juvela, Paavo Alku, Yu-Huai Peng, et al.

 

Abstract:
The ASVspoof 2019 dataset is a comprehensive collection of bona fide and spoofed speech data designed to support the development and evaluation of anti-spoofing countermeasures in automatic speaker verification (ASV) systems. The dataset addresses the vulnerability of ASV systems to three primary types of spoofing attacks: replay, speech synthesis, and voice conversion. The dataset is divided into two scenarios: Logical Access (LA), which focuses on synthesized and converted speech, and Physical Access (PA), which focuses on replay attacks. The dataset includes various state-of-the-art spoofing techniques to provide a challenging test bed for anti-spoofing research.

 

Data Creation Method:
Collected and synthesized using various methods, including text-to-speech synthesis, voice conversion, and replay attacks.

 

Number of Speakers:

  • 107 speakers

Total Size:

  • Several thousand audio samples

Number of Real Samples:

  • 54,000 real (bona fide) samples

Number of Fake Samples:

  • 233,960 fake (spoofed) samples

Description of the Dataset:

  • Approximately 3-4 seconds per utterance

 

Extra Details:
The dataset includes logical access and physical access scenarios to cover different types of spoofing attacks. It provides a comprehensive benchmark for the development and testing of anti-spoofing measures.

 

Data Type:

  • Audio files

Average Length:

  • Approximately 3-4 seconds per utterance

Keywords:

  • Automatic speaker verification, countermeasure, anti-spoofing, presentation attack, text-to-speech synthesis, voice conversion, replay, ASVspoof challenge, biometrics, media forensics

When Published:

  • July 15, 2020

 

Annotation Process:
Research on anti-spoofing measures, development and testing of ASV systems, and comparative analysis of different spoofing attack types.

 

Usage Scenarios:
Logical Access Dataset (for synthetic and converted speech attacks), Physical Access Dataset (for replay attacks)

 

Miscellaneous Information:
The dataset provides a challenging test bed for anti-spoofing research and includes various state-of-the-art spoofing techniques.

 

Credits:
Datasets Used:

  • ASVspoof 2019

Speech Synthesis Models Referenced:

  • Various text-to-speech synthesis, voice conversion, and replay attack methods
Dataset Link


Main Paper Link


License Link


Last Accessed: 7/1/2024

NSF Award #2346473