An end-to-end spoofing countermeasure for automatic speaker verification using evolving recurrent neural networks

Research in anti-spoofing for automatic speaker verification has advanced considerably in the last three years. Antispoofing is a particularly difficult pattern classification problem since the characteristics of spoofed speech vary considerably and can never be predicted with any certainty in the wild. The design of features suited to the detection of unpredictable spoofing attacks is thus a staple of current research. End-to-end approaches to spoofing detection with exploit automatic feature learning have shown success and offer obvious appeal. This paper presents our efforts to develop such a system using recurrent neural networks and a particular algorithm known as neuroevolution of augmenting topologies (NEAT). Contributions include a new fitness function for network learning that not only results in better generalisation than the baseline system, but which also improves on raw performance by 22% relative when assessed using the ASVspoof 2017 database of bona fide speech and replay spoofing attacks. Results also show that mini-batch training helps to improve generalisation, a technique which could also be of benefit to other solutions to the spoofing detection problem.