Enhancing Audio Privacy with Representation Learning through Source Separation and Robust Adversarial Techniques

Abstract

The significant rise in employing smart sensors and machine learning approaches in acoustic monitoring comes with speech privacy concerns during the data transmission from local devices to remote servers. This thesis investigates the integration of source separation and adversarial learning into learning latent representation for audio privacy-preserving systems. The proposed model first removes partial speech information from the input audio recording, and the privacy-preserving features are learned from the filtered signal through adversarial learning. The model is evaluated in an ablation study with source separation and adversarial learning components. The results suggest that combining both source separation and adversarial learning is effective in concealing speech presence in latent space while not degrading the performance in the acoustic monitoring task.