dc.description.abstract | This paper presents a VLSI chip design for support vector machine (SVM) and GMM-Supervector (Gaussian Mixture Model-Supervector) based speaker verification.
In SVM-Based method, the proposed chip consists mainly of a speaker feature extraction (SFE) module, an SVM module, and a decision module. The SFE module performs autocorrelation analysis, linear predictive coefficient (LPC) extraction, and LPC to cepstrum conversion. The SVM module includes a Gaussian kernel unit and a scaling unit. The purpose of Gaussian kernel unit is to evaluate the kernel value of a test vector and a support vector first. Four Gaussian kernel parallel processing elements (GK-PEs) are design to process four support vectors simultaneously. Each GK-PE is designed by a pipeline fashion and capable of perform 2-norm and exponential operations. An enhanced CORDIC architecture is presented to calculate the exponential value. In addition to the Gaussian kernel unit, a scaling unit is also developed in the SVM module. The scaling unit is used to perform scaling multiplications and complete the remaining operations of SVM decision value evaluation. Finally, the decision module accumulates the frame scores generated by all the test frames, and then compare it with a threshold to see if the test utterance is spoken by the claimed speaker. This chip design is characterized by its high speed, capable of handling a large number of support vectors in SVM.
In GMM-Supervector method, the proposed chip consists mainly of a speaker feature extraction (SFE) module, a Gaussian mixture model (GMM) module, an MAP module, an SVM module. The GMM module can help us to compute the result of GMM quickly, and we propose a new MAP module, which contains numbers of parallel MAP-PE, each MAP-PE can help us calculate Gaussian mean values after adaptation quickly, thereby this paper enhance the speed of the overall system. | en_US |