SPEECHFAKES: Generalized Voice Anti-Spoofing and Voice Biometrics

Artificial speech can nowadays be flexibly generated with speech synthesis and there are methods to alter speaker identity. Quality of artificial speech is no longer robotic but has reached the limit where a listener may no longer hear difference of real and artificial speech. Artificial and modified speech are known to deteriorate the performance of automatic speaker verification (ASV) in the form of spoofing attacks, and in the future we may face with new forms of “speech deepfakes”. Speech anti-spoofing is the task of computer-based differentiation of human and artificial speech from audio waveforms. SPEECHFAKES addresses anti-spoofing with a special focus on improving generality across datasets and attacks to improve their fault tolerance and explainability. Part of the project is implemented in collaboration with international collaborators. The project contributes new machine learning based detectors and contributes to public-domain datasets to promote further research.