Federated learning for privacy-preserving speech recognition

Abstract

Speech signal contains rich information encompassing gender, accent, speaking environment, and other speaker characteristics. Meanwhile, deploying high-performance speech applications often requires a large amount of training speech data, which are often collected from end-users. Therefore, protecting data privacy becomes a rising concern when speech data are employed to deploy commercial speech applications. That motivates the rising interest in designing “federated learning” for voice assistants and mobile applications. This chapter will introduce recent advances in federated learning foundation algorithms and applications for speech recognition, and general acoustic processing. Furthermore, it will introduce how federated learning-based speech processing techniques (e.g., average gradient and teacher–student learning) would connect to some critical data protection guidelines and public regulations, such as European Union's General Data Protection Regulation (GDPR) and California Consumer Privacy Act (CCPA).