A Speech Obfuscation System to Preserve Data Privacy in 24-Hour Ambulatory Cough Monitoring

Audio analysis of cough sounds can provide objective measures of respiratory clinical features such as cough frequency. Audio-based 24-hour ambulatory cough monitoring systems currently lead the way in providing these objective measures across a range of respiratory diseases. However, to preserve data privacy in cough audio recordings, there is interest to remove any identifiable information contained within patient and third-party speech. In this study we employed real-life patient audio recordings from the VitaloJAK 24-hour ambulatory cough monitoring device. We developed an audio-based speech obfuscation system that specifically detects and obfuscates intelligible speech while retaining cough events. An algorithm was developed to detect vowel sounds since most intelligible information is contained here. The detection algorithm employed audio features including energy, spectral centroid and an adaptive voiced speech feature. The detected vowel sounds were obfuscated by replacing the original audio signal with a synthetic version generated using the original energy and pitch but without formants information. The system was designed using seven hours of audio recordings from seven different patients with respiratory disease. The system was then evaluated on five 24-hour real-life patient audio recordings (120 hours in total) which consisted of 21.6 hours of intelligible speech along with 3,376 coughs. The system obfuscated 99.3% (21.5 hours) of intelligible speech while retaining 99.6% (3,362) of coughs. This speech obfuscation system can preserve data privacy while using 24-hour ambulatory cough monitors. Furthermore, it can retain cough events and other aspects of 24-hour cough recordings which may be of clinical interest.