PhD Thesis Defense: Bhawana Chhaglani, Privacy-aware Audio Sensing Systems for Health and Wellness
Content
Speaker:
Abstract:
Sensors and computing have become deeply embedded in daily life, from smart assistants in our homes to smartphones in our hands to wearables on our bodies. This pervasiveness holds immense promise for healthier buildings, safer workplaces, and personalized well-being by converting everyday devices into collaborative human and environment health monitors. However, there are fundamental challenges in realizing this vision. Sensors generate vast streams of data that may leak sensitive information; direct monitoring of health-relevant phenomena requires expensive or intrusive instrumentation; and always-on sensing demands energy and computation beyond what resource-constrained devices can sustain. These shortcomings create a deployment and adoption gap that motivates my work.
In this thesis, I present privacy-aware audio sensing systems for monitoring environmental and human health. By leveraging everyday acoustic signals and lightweight machine learning techniques, these systems enable deployable and non-invasive sensing solutions for applications including ventilation monitoring, airborne risk estimation, speech health assessment, and privacy-preserving audio inference. Across these applications, I focus on developing sensing methodologies that preserve privacy while maintaining strong utility, enabling trustworthy health sensing in real-world environments.
First, I present FlowSense, a non-invasive acoustic sensing system for monitoring building ventilation. By capturing low-frequency airflow sounds from ventilation systems, FlowSense predicts vent status and airflow rates using privacy-preserving signal processing techniques and on-device machine learning models. The system enables real-time assessment of ventilation adequacy using commodity microphones without requiring specialized infrastructure.
Next, I present AeroSense, an ambient acoustic sensing system designed to estimate aerosol emissions generated by everyday human activities. AeroSense uses non-reconstructible audio features to detect and localize activities such as speaking, coughing, and sneezing, while also inferring contextual factors including mask usage and interpersonal distance. By combining these parameters, AeroSense estimates airborne transmission risk and provides actionable insights for maintaining healthier indoor environments.
Next, I present DysSense, a privacy-aware and interpretable speech health assessment system for fine-grained monitoring of neurological speech impairments. DysSense introduces subsystem-aware biomarker modeling across respiration, phonation, articulation, and prosody, combined with selective speech suppression techniques that preserve clinically relevant biomarkers while minimizing intelligible speech leakage. I further propose a reference-free metric for online speech leakage assessment without requiring ground-truth transcripts.
Finally, to address broader privacy concerns in audio sensing, I introduce FeatureSense, a privacy-aware audio sensing framework that systematically evaluates and mitigates speech and speaker-related privacy leakage from commonly used acoustic features. FeatureSense proposes a general-purpose library of privacy-preserving features and adaptive feature selection strategies that balance utility, privacy, and computational efficiency across diverse sensing applications.
Collectively, these systems demonstrate how privacy-aware AI-driven sensing can transform everyday devices into trustworthy collaborators that sense, reason, and act for our well-being. By addressing challenges in privacy, deployability, and robustness, this thesis advances the foundations of trustworthy audio sensing for environmental and human health applications.
Advisor:
Prashant Shenoy