A Hybrid Wav2Vec2 and LFCC-CNN Framework for Voice Spoofing Detection

Abstract

Anti-spoofing systems are important for protecting security-sensitive applications such as financial services, smart assistants, biometric authentication, and access control systems. However, the increasing use of text-to-speech (TTS) and voice conversion (VC) technologies has made it easier to generate spoofed speech that closely resembles genuine human speech. Therefore, developing reliable methods for distinguishing genuine and spoofed speech remains an important challenge. A hybrid voice spoofing detection framework is proposed in this work by fusing Wav2Vec2 representations with Linear Frequency Cepstral Coefficient (LFCC) features. Wav2Vec2 extracts contextual representations directly from raw speech, while LFCC features capture fine-grained spectral characteristics associated with artifacts introduced by speech synthesis and voice conversion. The LFCC features are encoded using a lightweight Convolutional Neural Network (CNN), and the resulting 256-dimensional representation is fused with the 768-dimensional Wav2Vec2 embedding to form a 1024-dimensional feature representation. The fused representation is subsequently processed by a fully connected classification network to distinguish bona fide and spoofed speech. The proposed framework is evaluated on the ASVspoof2019 Logical Access (LA) dataset using standard anti-spoofing evaluation metrics. Experimental results show that the proposed hybrid model achieves an accuracy of 98.95% and an Equal Error Rate (EER) of 1.02%. An ablation study comparing the individual Wav2Vec2 and LFCC-CNN branches with the proposed hybrid model shows the benefit of combining contextual and spectral representations for voice spoofing detection.