From media fears to research reality: How ready are countermeasures against speaker verification spoofing?

The subject of biometrics spoofing has received more attention over the last few years within the research community. For speech technology spoofing attacks [1] relate to replay attacks [2, 6] and a category sometimes referred to as synthetic voice attacks, including attacks with artificial signals such as voice conversion and speech synthesis [3, 4, 5]. Some significant initiatives came with the ASVspoof2015 evaluation focused on spoofing with synthetic voice [3] and more recently ASVspoof2017, an evaluation focusing on replay detection [6]. We present some results about the latest development on voice anti-spoofing, most of them derived from work undertaken during the H2020 OCTAVE project [7]. Results on synthetic voice detection [Figure 2] are presented with synthetic audio from systems recently mentioned in press coverage [Figure 1]. More results are presented about the recent development of replay detection systems within the Octave project and for the ASVspoof2017 challenge. Finally, we discuss the challenges ahead, with results showing the difficulty of building up representative databases and further results showing adverse effects of codec and noise.