Parallel implementation of a VQ-based text-independent speaker identification

Abstract

This study presents parallel implementation of a vector quantization (VQ) based text-independent speaker identification system that uses Mel-frequency cepstrum coefficients (MFCC) for feature extraction, Linde-BuzoGray (LBG) VQ algorithm for pattern matching and Euclidean distance for match score calculation. Comparing meaningful characteristics of voice samples and matching them with similar ones requires large amount of transformations and comparisons, which result in large memory usage and disk access. When the cost of computations is considered, it states the main motivation for a parallel speaker identification implementation, where the parallelism is achieved using domain decomposition. In this paper, we present a set of experiments using the YOHO speaker corpus and observe the effects of several parameters as VQ size, number of MFCC filter banks and threshold value. First we focus on the serial algorithm and improve the algorithm to give the best success rates and provide a strong base for parallel implementation, where a clear performance improvement on speedup is obtained.