Acoustic Sensing for Assistive Living: Investigating ML Models to Address Privacy and Data Collection Challenges

Assistive living technologies have advanced significantly, utilising wearable sensors and other smart devices. While wearables are effective in monitoring human activity, they may become impractical when the elderly are resistant to them. In contrast, audio sensing can passively monitor daily routines without the burden of managing wearable devices. This research is divided into three parts: the first focuses on an unsupervised application of acoustic sensing, the second distinguishes data collection and application challenges in an acoustic environment, and the third presents a proposed pipeline that can solve privacy concerns regarding speech in an acoustic environment. Audio as a modality is rich with contextual information and can be very useful in sensing activities in a home environment. Because of the nature of an audio-sensing environment, there are some restrictions we have to keep in mind in terms of data collection and privacy. Firstly, having a large amount of labelled data is not possible. Therefore, this thesis's first part focuses on an unsupervised approach to detect changes in daily routine in an acoustic environment with a 14% improvement in the F1 score compared to the previous baseline. The second part of our research sets the challenges of any acoustic home environment regarding data collection and speech privacy, showcasing the solution deployed in the industry. Finalising this thesis with the creation of a three-stage methodology on developing a privacy pervasive system, which is a system that ensures privacy is maintained at all stages of data processing, regarding speech audio data and the possibility of having a passive audio sensing system that discards all speech data and then having the possibility to apply any audio sensing techniques without compromising privacy. The proposed methodology achieved this by eliminating all private conversations from the acoustic signal, resulting in a minor reduction of 2-13% in the F1 score for acoustic activity detection across various datasets.