Modified Group Delay Cepstral Coefficients for Voice Liveness Detection
Shrishti Singh, Kuldeep Khoria, Hemant A. Patil · 2021 29th European Signal Processing Conference (EUSIPCO) · 2021
Liveness detection has advanced for many biometrics, such as face, iris, hand geometry, etc. However, less emphasis is given to liveness detection for voice biometrics, i.e., Voice Liveness Detection (VLD). Pop noise is produced due to the spontaneous breathing while uttering a certain class of phonemes (such as fricative, affricates, plossive, nasal, etc.) which has low frequency characteristics. In this paper, phase-based information for analysis of pop noise is used as a feature for VLD task. We have use modified group delay function based cepstral coefficients (MGDCC) as a feature for VLD task. Experiments are performed using two different types of classifiers, i.e., Gaussian Mixture Model (GMM) and Convolutional Neaural Network (CNN). Our results indicate an overall improvement in accuracy by 17.16% and 1.38% for MGDCC-CNN and MGDCC-GMM systems respectively.