On-Device Speech Filtering for Privacy-Preserving Acoustic Activity Recognition

Acoustic sensing has become increasingly prevalent for mobile and ambient devices for applications such as human activity recognition, health monitoring, and environmental sensing. These approaches develop audio featurization techniques to enable machine learning-based inferences while offering privacy by preventing speech reconstruction. However, recent work [2] has shown that such methods are still vulnerable to speech content recovery when fine-tuned automatic speech recognition (ASR) models are applied. Here, we demonstrate the broad applicability of on-device speech filtering using a detect-and-remove method for acoustic sensing tasks, significantly reducing privacy risks in revealing speech content. Additionally, we introduce an interactive tool to experiment with audio featurization methods, aiding the development of privacy-preserving applications.