A novel multiclass classification based approach for playback attack detection in speaker verification systems

Spoofing detection in automatic speaker verification (ASV) systems is typically handled as a binary classification approach. In this paper, we propose a novel approach to address this problem using a multi-class classification approach. Each audio sample is tagged on the basis of the source of the signal. Spoof class samples are divided according to corresponding recording devices which were used during recording of the genuine speaker’s voice to be later used for implementing playback attack. Three different multiclass based approaches proposed in this work are evaluated on ASVspoof 2017 v2.0 dataset. The performance of these systems is tested on conventional and deep classifier systems using both handcrafted features and spectrographic representations of audio. Results suggest the potential of the proposed multiclass classification based approach in comparison to binary classification, specifically in deep learning scenario.