Investigating the use of scattering coefficients for replay attack detection

Abstract

Widespread adoption of speaker verification for security relies on the existence of effective anti-spoofing countermeasures. This paper presents a countermeasure based on spectral features to detect replay spoofing attacks on automatic speaker verification systems. In particular, the use of hierarchical scattering decomposition coefficients and inverse- mel frequency cepstral coefficients are explored. Our best system achieved a relative improvement of around 70% in terms of equal error rate on the development set and 20% on the evaluation set, when compared to the baseline on the ASVspoof 2017 database. In addition, we show that features with a shorter window can be beneficial to detecting replayed speech, in contrast to speech synthesis and voice conversion attack.