A Survey on Linear Prediction Residual based Replay Attack Detection System
Madhusudan Singh · 2021
It is well known fact that modern automatic speaker verification (ASV) system is vulnerable to replay attacks, and this vulnerability has increased, in light of rapid progress in recording and playback device manufacturing technologies. Prior studies widely explored spectral features, modulation features and phase features in developing replay detection systems for ASV protection. In this order, focus on excitation source features has been comparatively very less. The linear prediction (LP) residual represents excitation source information implicitly. This paper provides a detailed review on prior LP residual derived excitation source features exploited for replay detection task. As outcome, we inferred that exploring LP residual as excitation source features using suitable signal processing algorithms may provide better solutions than existing. Further, these excitation source features are complementary to those of well performing spectral/modulation and phase features, and can be used in combination in order to obtained better solutions close to generalization. After this, the usefulness of LP residual information for replay detection task has been discussed from future perspective. It is concluded that capturing harmonic related excitation source information from LP residual signal may provide better results in replay speech detection context.