Speech privacy-aware acquisition of acoustic information based on deep learning algorithms

Abstract:
Acoustic sensors could potentially add relevant information about the status and environment of vehicles and, hence, be used for increased safety and comfort in automotive applications. However, apart from technical challenges, employing exterior microphones requires addressing possible privacy issues, because it cannot be excluded that intelligible voices of persons in the proximity of the vehicle are recorded without their explicit consent, which would violate existing legal regulations. This contribution therefore presents a system which automatically detects speech activity in a recorded audio stream. This information can then be used to process the audio signal accordingly, for example by modifications that protect the identity of the talkers while maintaining the information relevant for the safety of the vehicle (e.g., approaching sirens, shouts for help, changes in road surface etc.). The validations show that a reliable detection of speech activity can be achieved even in adverse urban noise conditions, indicating that privacy-related issues can be overcome with appropriate algorithmic solutions.