Privacy Preserving Continuous Speech Recording using Throat Microphones

A prerequisite for field-research on audio data are privacy preserving recordings that exclusively contain the target speaker who gave consent. For this purpose, we investigated the potential of a simple but robust wearable technology consisting of three parts: first, a standard air-conduction microphone providing the necessary audio quality for speech analysis; second, a throat microphone used as a speech activity filter; third, a custom ESP32 based recording device enabling on-device real-time processing. The system was evaluated in two challenging free discussion settings with two and four participants each (total N=16). Results from manual annotations show an Equal Error Rate of M=23.4-29.69 %. Based on simple instructions, our participants managed to maintain a False Acceptance Rate below 5 % while recording more than half of their utterances.