Hardware Architectures for Embedded Speaker Recognition Applications: A Survey

Authentication technologies based on biometrics, such as speaker recognition, are attracting more and more interest thanks to the elevated level of security offered by these technologies. Despite offering many advantages, such as remote use and low vulnerability, speaker recognition applications are constrained by the heavy computational effort and the hard real-time constraints. When such applications are run on an embedded platform, the problem becomes more challenging, as additional constraints inherent to this specific domain are added. In the literature, different hardware architectures were used/designed for implementing a process with a focus on a given particular metric. In this article, we give a survey of the state-of-the-art works on implementations of embedded speaker recognition applications. Our aim is to provide an overview of the different approaches dealing with acceleration techniques oriented towards speaker and speech recognition applications and attempt to identify the past, current, and future research trends in the area. Indeed, on the one hand, many flexible solutions were implemented, using either General Purpose Processors or Digital Signal Processors. In general, these types of solutions suffer from low area and energy efficiency. On the other hand, high-performance solutions were implemented on Application Specific Integrated Circuits or Field Programmable Gate Arrays but at the expense of flexibility. Based on the available results, we compare the application requirements vis-à-vis the performance achieved by the systems. This leads to the projection of new research trends that can be undertaken in the future.