Development of a Synthetic Database for Compact Neural Network Classification of Acoustic Scenes in Dementia Care Environments

This paper focuses on automatic detection and classification of sounds occurring in dementia care facilities for monitoring a resident's safety and wellbeing. While there has been significant advances the field of domestic audio classification within the recent years and several audio databases exist, these have not been designed for dementia care environments and can be limited in terms of the amount of information they provide, such as the exact location of the sound sources, and the associated noise levels. This work details our approach to generating a synthetic database of sound scenes and events that is carefully curated to reflect a typical real-world dementia care environment. This includes background noise and room impulse responses based on a typical one-bedroom apartment (Hebrew SeniorLife Facility). The database contains clean and noisy excerpts from 11 classes with duration of 5-seconds and sampling rate of 16 kHz. Using this database, we also explore further development of a series compact neural network architecture through our baseline model which utilizes Continuous Wavelet Transform scalograms as features to the AlexNet. Our compact, MAlexNet-40 approach has achieved a 15x reduction in network size, and an improvement of about 3% on the weighted F1-score when compared to the traditional AlexNet model.