Spoofing Detection for Personal Voice Assistants

Personal Voice Assistants (PVAs) are common acoustic sensing systems that are used as a speech-based controller for critical systems making them vulnerable to speech spoofing attacks. Prior research has focused on the discrimination of genuine and spoofed speech for applications with large population speaker verification and challenges such as ASVspoof have advanced this work over the last few years. In this paper, we consider spoofing detection in a PVA setting where the number of household users is small. We show that when pre-trained models are adapted to household users, spoofing detection is improved. Furthermore, we demonstrate that adaptation is still effective in realistic scenarios where only genuine speech of household users is available but the generation of spoofed speech samples for household users is undesirable.