Year: 2020
Type: dissertation
Abstract: Recently, spoofing attacks, e.g.replay, speech synthesis, voice conversion, have become a serious problem against automatic speaker verification (ASV) system.One of countermeasures assuming multi-channel recording has been proposed since smartphones and smart speakers have been popular and multiple microphones have been installed in them.This method focuses on the fact that different phonemes are emitted from different locations within the human vocal tract system, where generalized cross-correlation (GCC) that estimates a position of sound source are used for phonemes localization.While the method has situation dependencies, GCC motivates us to consider other methods based on loudspeakerspecific characteristics.We propose a replay attack detection method using a difference between genuine speech and replayed one by using the peak value of GCC.In experiments, the proposed methods are confirmed to provide low error rates without environmental restrictions.In addition, score fusion systems with other acoustic feature-based methods are demonstrated to improve robustness compared to conventional methods.1.
Author: ヤグチ, リョウヤ
Language: Japanese
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7208359778
