Linear Prediction Residual-Based Constant-Q Cepstral Coefficients for Replay Attack Detection
Khomdet Phapatanaburi, Prawit Buayai, Mongkol Kupimai, Teerapon Yodrot · 2020
This paper proposes a new linear prediction residual-based constant-Q cepstral coefficients (LPR-CQCC) feature for replay attack detection (RAD). The main contribution of the proposed feature is to modify conventional constant-Q cepstral coefficients (CQCC) using linear prediction residual (LPR) signal instead of the original/raw speech signal. Since the LPR signal has a distortion obtained from playback devices, leading to the a difference between actual and replayed speech signal, it is expected that extracting the proposed feature based on LPR signal may provide a promising result for RAD task. To further improve the detection performance, LPR-CQCC was also combined with original CQCC and Gammatone-scale relative phase (Gammatone-scale RP) in order to fuse the complementary advantages based on different systems at score-level. Based on Gaussian mixture model-based classifier, the results on ASVspoof 2017 database version 1 exhibited that LPR-CQCC outperformed baseline CQCC on the development set and was very close to CQCC on the evaluation set. Moreover, the score combination of the proposed feature and CQCC/Gammatone-scale RP performed better than the system which uses individual features.