Pseudo-Voice Processing: Adversarial Robustness and Spectral Consistency in GAN-VAE for Synthetic Audio Security

This paper presents a comprehensive framework for pseudo-voice processing integrating GAN-VAE hybrid architecture and convolutional neural networks (CNNs), addressing security challenges in AI-generated speech. The system combines generative adversarial networks (GANs) with variational autoencoders (VAEs) to enable high-fidelity pseudo-voice synthesis while enhancing synthetic audio detection. Key innovations include a novel spectral consistency loss for disentangled speaker/prosody control, adversarial stability constraints for robustness, and a CNN classifier optimized via openSMILE preprocessing. Experimental results demonstrate 92.3% accuracy and 0.89 pseudo-voice sensitivity, outperforming baselines. The framework advances security in voice authentication by detecting synthetic artifacts, such as spectral irregularities and over-regularized prosody.