Thus, false content may be employed for impersonation, disinformation, or fraudulent access to automated services, and can be forged in many different ways, from very simple signal manipulation to the application of recent machine learning techniques, which have considerably increased the plausibility and damaging capability of the generated content. This talk will focus on the detection of spoofed and fake speech signals.

Two different approaches have been proposed to detect these attacks. On one hand, passive solutions try to directly determine whether a given speech utterance is genuine or spoofed/fake by a deep analysis of the signal itself. This is the most common approach and has been extensively studied over the last decade, boosted by a number of challenges dealing with different types of spoofed or fake speech signals. On the other hand, proactive solutions require the collaboration of the digital content provider, who must watermark the speech signal in order to allow the detection of its synthetic origin, thus avoiding any malicious use of the generated speech. Also, we must not forget that the generation of fake contents follows the dynamics of a “cat-and-mouse game,” so it is very important to understand how attacks may be forged in order to be ready to combat them. In this talk, we will review how all these topics have evolved over the last years and study the most recent, state-of-the-art techniques. We will also go through a hands-on exercise to put some of the talk’s concepts into practice.