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

Abstract

Automatic user identification is an indispensable part of today's computer based applications. Passwords and keys are common solutions due to their ease of implementation and low cost, but these solutions also contain the risk of being forgotten, stolen, or being used by unauthorized users, therefore security professionals are working on biometric solutions that are based on human specific characteristics. Voice is a popular biometric as it can be easily collected and digitalized by a microphone set or by a phone. In this study a text-independent speaker identification system is presented. Mel-Frequency Cepstrum Coefficients are used in feature extraction, Linde-Buzo-Gray vector quantization is used in modeling these features, and measuring the similarity of models is achieved by using Euclidean distance metric. Comparing meaningful characteristics of voice samples requires a significant amount of transformations and calculations, to share this load to a cluster system instead of using a serial machine, a parallel text-independent speaker identification system is implemented, and clear performance improvements are observed. Our parallel speaker recognition system achieves a speed up about 13.8 compared with its serial implementation in the case of using 16 processing elements to identify a corpus of 100 speakers.