A method for preserving privacy during audio recordings by filtering speech

Continuous monitoring of sensor data from wearable devices such as smartwatches has been a hot topic recently and has many practical applications, particularly towards health monitoring. Both industry and academia have been using sensors such as accelerometers, gyroscopes and pulse oximeters, to infer and monitor, among other activities, physical activity, heart rate, sleep and walking. However, one sensor that tends to go unused, despite it's vast utility, is the microphone. The microphone has the ability to pick up potentially useful sounds such as coughing, wheezing, throat clearing, sneezing and more. The reason it often goes unused is because of the privacy concerns associated with continuously recording audio. Speech of the user and others around the user can be picked up by the microphone, which is generally considered highly invasive. In this paper, we propose a method for continuously recording audio while preserving privacy by filtering sensitive signals such as speech. We explain how this method can be applied to create classifiers and deployed for in-the-wild studies. We use cough detection as an example application and show that the proposed speech filtering method does not degrade the classification accuracy when detecting coughs.