Sound is a powerful source of information. By training algorithms to identify different sound patterns, we can learn what a person is doing, like cooking, vacuuming, or washing dishes. However, using sound for this purpose raises privacy concerns since microphones can pick up sensitive information.

To address this, researchers at Carnegie Mellon University created a filter called Kirigami that can detect and remove human speech segments collected by audio sensors before they are used for identifying activities. This helps protect people’s privacy while still allowing for audio sensing.

Existing privacy techniques in audio sensing involve altering data or training computers to ignore certain frequencies, but advances in generative AI have made this less effective. Kirigami helps prevent AI models from accessing speech snippets in processed audio data.

This lightweight filter can run on small microcontrollers and remove speech content to ensure sensitive data stays on the device. It uses a binary classifier to determine if there is speech in the audio and allows for configurable thresholds to balance privacy and utility.

Kirigami is a promising solution for developers looking to balance privacy and functionality in audio sensing devices. It has potential applications in areas like health monitoring, environmental sensing, and activity recognition. As smart home technology grows, developers can customize Kirigami to meet their specific privacy needs.