Novel Amplitude Weighted Frequency Modulation Features for Replay Spoof Detection

Abstract

In this paper, we propose the use of Amplitude Modulation and Frequency Modulation (AM-FM) features for replay detection task. In AM-FM signal, AM component is known to be severely affected by noise (in this case, due to replay mechanism) which is exploited in proposed feature extraction. In particular, we explore this damage in AM component to corresponding instantaneous frequency. Thus, the novelty of proposed Amplitude Weighted Frequency Cosine Coefficients (AWFCC) feature set is the use of frequency components along with weighted amplitude components that are degraded due to replay noise. The AWFCC features have the information of AM-FM and hence, gave discriminatory information in the spectral characteristics. The experiments were performed on ASVspoof 2017 Challenge database using AWFCC feature set that gave an Equal Error Rate (EER) of 6.37 % (dev) and 11.72 % (eval) set. We have compared our results with CQCC, LFCC and MFCC using Gaussian Mixture Model (GMM) classifier and found that proposed AWFCC feature set performed better than the other feature sets on both dev and eval datasets. In addition, we have used score-level fusion of CQCC and AWFCC to obtain a lower EER of 3.60 % and 11.22 % on dev and eval set.