Speech Obfuscation in Mel Spectra that Allows for Centralised Annotation and Classification of Sound Events

Nowadays, computerised Sound Event Classification (SEC) aids in several applications, e.g. monitoring domestic events in smart homes. SEC model development typically requires data collected from a diverse set of remote locations. However, this data could disclose sensitive information about uttered speech that might have been present during the acquisition. In this work, three data preprocessing techniques are investigated that obstruct recognising semantics in speech, but retain the required information in the data for annotating sound events and SEC model development. At the remote location, the data are first preprocessed before transferring to a central place. At the central location, speech should not be interpretable anymore, while still having the opportunity to annotate data with relevant sound event labels. For this purpose, starting from a log-mel representation of the sound signals, three speech obfuscation techniques are assessed: 1) calculating a moving average of the log-mel spectra, 2) sampling a few of the most energetic log-mel spectra and 3) shredding the log-mel spectra. Both intelligibility and SEC experiments were carried out. All considered techniques proved effective in obfuscating speech, while still allowing SEC. For stationary sound events, calculating the moving average of the log-mel spectra is recommended, as well as shredding the log-mel spectra. For impulsive sound events, sampling a few of the most energetic log-mel spectra is recommended.