Voice livness detection based on pop-noise detector with phoneme information for speaker verification

Abstract

This paper proposes a pop-noise detector using phoneme information for a voice liveness detection (VLD) framework. In recent years, spoofing attacks (e.g., reply, speech synthesis, and voice conversion) have become a serious problem against speaker verification systems. Some techniques have been proposed to protect the speaker verification systems from these spoofing attacks. The VLD framework has been proposed as one of fundamental solutions. The VLD framework identifies that an input sample is uttered by an actual human or played by a loudspeaker. To realize the VLD framework, pop-noise detection methods have been proposed and these methods perform well as the VLD module. However, since pop-noise is a common distortion in speech that occurs when a speaker’s breath reaches a microphone, the phenomenon of pop-noise is able to be occurred by winds or attackers arbitrary. It is one problem of the pop-noise detection methods. In order to improve the robustness of the pop-noise detection methods, this paper proposes a pop-noise detector using phoneme information as an evidence of an actual human. From the experimental results, the proposed method increases the robustness of the VLD against spoofing attacks.

Securing Liveness Detection for Voice Authentication via Pop Noises

“…We confirm the effectiveness of our system against replay attacks and impersonation attacks by comparing it with the baseline (VLD) [18] and the conference version of VoicePop [1]. In [18], Sayaka Shiota et al proposed the pop noise detector combined with the phoneme information to detect the existence of pop noises, but the replayed samples were easily recognized as legitimate samples under their proposed algorithm.…”

“…We confirm the effectiveness of our system against replay attacks and impersonation attacks by comparing it with the baseline (VLD) [18] and the conference version of VoicePop [1]. In [18], Sayaka Shiota et al proposed the pop noise detector combined with the phoneme information to detect the existence of pop noises, but the replayed samples were easily recognized as legitimate samples under their proposed algorithm. However, VLD does not consider using the characteristics of the pop noise for further classification, nor does it consider the impersonation attack when the adversary replays the audio and mimics breathing at the √ Earise Al-101 √ This also shows that the combination of pop noise and its airflow pressure can improve the detection rate of the pop noise-only feature.…”

“…Sayaka Shiota et al [17] proposed the pop noise detector, which combines the single-and the doublechannel to detect pop noise. They further incorporated the phoneme information for pop noise detection in [18]. However, their studies rely on the specific microphone model and cannot perform well when applied to mobile devices.…”