Liveness Detection – Automatic Classification of Spontaneous and Pre-recorded Speech for Biometric Applications
Cristian-Teodor Neamtu, Şerban Mihalache, Dragoş Burileanu · 2023
Biometric systems based on speaker recognition are susceptible to attacks consisting of an impostor (unauthorized user) presenting a pre-recorded speech sample obtained from a genuine user. This type of liveness detection is called a replay attack. This paper proposes an automatic speaker verification system to classify spontaneous speech (from a legitimate user) and pre-recorded speech samples (leveraged by an unauthorized user) using classical machine learning algorithms and deep neural networks. The system was evaluated on two corpora: ASVspoof 2017 and ASVspoof 2019. These datasets contain genuine and spoofed utterances, recorded in realistic conditions, and pre-split into training, development (dev), and evaluation (eval) subsets. Final performance for the best proposed systems reaches an equal error rate (EER) of 18.9% (dev) and 26.1% (eval) for the ASVspoof 2017 dataset, and 10.2% (dev) and 15.2% (eval) for ASVspoof 2019.