Room identification with Personal Voice Assistants

Abstract

Personal Voice Assistants (PVAs) are used to interact with digital environments and computer systems using speech. In this work we describe how to identify the room in which the speaker is located. Only the audio signal is used for identification without using any other sensor input. We use the output of existing trained models for speaker identification in combination with a Support Vector Machine (SVM) to perform room identification. This method allows us to re-use existing elements of PVA eco-systems and an intensive training phase is not re- quired. In our evaluation rooms can be identified with almost 90 percent accuracy. Room identification might be used as additional security mechanism and the work shows that speech signals recorded by PVAs can also leak additional information.