Speaker identification from extracted features of selective energized voice signal

Abstract

Voice is both physiological and behavioral trait of a human. Hence, the voice based identification system is convenient for both on-site and remote ways. Despite having some important aspects, the speaker based identification system is one of the most challenging fields in the biometric-based authentication system. These challenges occur especially due to environmental or channel based noise, lack of training samples, and the difference between training and testing environment. However, different noise elimination techniques stand for the different type of noises. And, finding the noise type from a voice signal is not an easy task. This research has focused on two important aspects of these challenges: firstly, high energetic portion of a voice signal has been focused, as high energy portion of signal is less contaminated by noise, and secondly, speaker identification has been carried out with less trained data for the scenario where multi conditional training data is not available. To achieve this aim, the high energetic signal portion has been selected to extract voice features and hence Linear Prediction has been applied to the selected portion of dynamic model orders. Dynamic Model Order (DMO) approach is determined by the number of local maxima of the selected signal portions which are followed by the feature extraction step is performed using Cepstral Coefficient (CC) analysis. This approach shows a significant accuracy result than the stateof- art technology.