A privacy stack for speech-based AI in digital health

Speech is an accessible and information-rich clinical signal, but its diagnostic value is deeply entwined with biometric identity. Empirical and perceptual evidence shows that current anonymization methods cannot fully remove identity cues and often introduces disorder-dependent artifacts. These limitations reveal a need for a system-level approach. We propose a privacy stack that reconceives privacy as identity uncertainty, achieved through coordinated signal diversification, leakage-resistant model design, and privacy-aware infrastructure. This framework outlines a path toward clinically meaningful, equitable, and trustworthy speech-based artificial intelligence systems.