Voice Spoofing Classification Using Residual Bidirectional Long Short Term Memory

Voice spoofing attacks are a major security concern for speech-based biometric systems. Detection and classification of spoofed voice are essential steps for preventing unauthorized accesses. This study proposes a novel approach to voice spoofing classification using a Residual Bidirectional Long Short Term Memory (R-BLSTM) network. The goal is to enhance the accuracy and robustness of voice spoofing detection using the power of deep learning and residual connections. The current proposed approach based on bidirectional LSTM with residual connections is designed to capture long-range dependencies and latent characteristics of speech signals. Experimental evidence that the R-BLSTM model is superior to classic ML techniques is also demonstrated by observing an accuracy of 95.6% on the ASVspoof 2019 collection. The designed system can be further utilized for enriching the security of speech-based biometrics modalities and making anti-voice spoofing attacks ineffective.