Area Efficient CORDIC based SVM for Speaker Verification System

this paper presents the implementation of a support vector machine (SVM) for speaker verification system. The proposed system includes a Gaussian kernel unit and a scaling unit. The proposed system can be used within a speaker verification system for checking the validity of a speaker. A Gaussian kernel processing element (GK-PEs) and an exponential processing element are included in the support vector machine. The Gaussian kernel processing element is associated with the evaluation of kernel value of the test vector and the supporting vector. The Gaussian kernel processing element is designed to process four supporting vectors simultaneously. The GK-PE is designed in a pipeline fashion and therefore can perform two-norm and exponential operations. An enhanced CORDIC architecture is used inside the SVM in order to calculate the exponential value.. The SVM decision value evaluation is performed by the Scaling Unit. To be used inside a speaker verification system an additional module called Speaker Feature Extraction (SFE) module is also needed. The SFE module is responsible for performing autocorrelation analysis, linear predictive coefficient (LPC) extraction, and LPC-to-cepstrum conversion. The frame scores that are generated for the test frames are then accumulated by the decision module and are then compared with a threshold to check if the test utterance is spoken by the claimed speaker.