A CUDA based Parallel Implementation of Speaker Verification System

ABSTRACT: Speaker Verification (SV) is a type of speaker recognition that validates the identity of a claimed person by his/her voice. Training the models from large speech data requires a significant amount of memory and computational load. In this thesis we present a parallel implementation of speaker verification system based on Gaussian Mixture Modeling – Universal Background Modeling (GMM – UBM) designed for many-core architecture of NVIDIA’s Graphics Processing Units (GPU) using CUDA single instruction multiple threads (SIMT) model. CUDA implementation of these algorithms is designed in such a way that the speed of computation of the algorithm increases with number of GPU cores. In our experiments we have achieved 30 times speedup for k-means clustering and 65 times speedup for Expectation Maximization (EM) for an input of about 350K frames of 16 dimensions and 1024-2048 mixtures on GeForce GTX 570 (NVIDIA Fermi Series) with 480 cores when compared to a single threaded implementation on the traditional CPU. Keywords: Speaker Verification, Gaussian Mixture Models, Parallel Computing, Compute Unified Device Architecture, General-purpose computing on graphics processing units. ……………………………………………………………………………………………………………………………………………………………………………………………………………………