Privacy-Preserving Speech Recognition System—A Conceptual Model

A lot of user speech data is being accumulated as Automatic Speech Recognition (ASR) are integrated into devices to improve these systems. Privacy protection in speech data has drawn increased attention since the General Data Protection Regulation (GDPR) was implemented in the EU. As such voice data contains Personal Information (PI) which may jeopardize users’ right to privacy. These devices also capture voice passively even when the user is not interacting with the application. This poses a serious threat to the entity's sensitive information. This initiated the requirement of safety precautions for the use of voice data. The goal of this study is to discover methods for maintaining voice message/command security without changing the required content to maintain the data utility while safeguarding users’ privacy. We suggest a model, Advanced Automatic Speech Recognition (AASR) that segregates the confidential information of the user from the data that is required by the device/application. This is implemented in a phased manner where Support Vector Machine (SVM) is used in the first phase and adversarial noise in the second phase. The conceptual model emphasizes privacy by SVM removing most of the sensitive data and adversarial noise masking the remaining sensitive information. We also deduce any potential future propositions of the model.