FeatureSense: Protecting speaker attributes in always-on audio sensing systems

The FeatureSense system targets always-on audio sensing and provides a privacy solution that balances sensing utility with protection of speaker attributes such as age and gender; the study also notes that these attributes can still be inferred after masking speech content. The system provides general-purpose privacy-aware audio features as an open-source library and uses task-adaptive feature selection to trade off privacy, utility and cost. In evaluations across multiple sensing tasks, its effectiveness at protecting user-specific privacy is 60.6% higher than existing privacy techniques.