A Low Power Look Up Table-Free Gaussian Mixture Model Based Speaker Classifier

In this thesis an ASIC design of an hardware GMM-Based Speaker Classifier is presented.The Classifier is a fundamental component of a Speaker Identification system, that is ableto associate an unknown incoming speech signal to its unknown speaker, which is part of apreviously modeled group of speakers. The design flow follows a software-to-hardware approach,since the whole system is firstly implemented in Matlab, then increasingly transformed from high-level to machine-level until its hardware description. Innovative techniques that avoid any memory accesses to perform hardware exponentials and logarithms are presented. All the computations are executed on-chip and this gives extra performances and extra security to the system. Thanks to its low power demand, it is suitable to be integrated in many IoT devices to personalize the user experience without compromising the power budget of the device