Deep Convolutional Neural Network for Voice Liveness Detection

In this work, we present the system to detect the liveness by identifying the pop noise in the voice signal in order to avoid the security breach of ASV systems. Pop noise is created due to spontaneous breathing while uttering a certain phonemes, and it has low-frequency characteristics. Given the low-frequency characteristics of the pop noise, we have used the short-time Fourier transform (STFT) with low-frequency contents (0-40 Hz) as a feature set along with a convolutional neural network as a classifier. The experiments are performed using the recently released POp noise COrpus (POCO) dataset. We have considered the approach given in the original POCO dataset paper as a baseline and compared the results with the proposed architecture. The performance of the proposed architecture is measured using 10-fold cross-validation and the customized disjoint partition of the dataset. It is observed that the proposed architecture shows an improvement in accuracy for voice liveness detection in both cases. In particular, the proposed architecture obtained 19.86% and 18.22% absolute improvement in accuracy for 10-fold cross-validation and customized data partition, respectively, as compared to the baseline.