Dialect Recognition System Using Excitation Source Features

Abstract

This paper focuses on building an automatic dialect recognition system using excitation source features. Every spoken unit represents the unique articulatory configuration of the excitation source and the vocal tract system. This paper emphasis on exploring source information to capture dialectal cues over vocal tract information. Epochs representing the instants of maximum excitation of the vocal tract at the closure are used as source features. Additionally, strength and slope of epochs and instantaneous frequency features are extracted from zero frequency filtered signal. Further, 13 cepstral coefficients are derived from the LP residual to prepare feature vector. Two dialect datasets such as Kannada dataset with five prominent dialects and English dataset with nine dialects are used for evaluation of the significances explored features. Classification experiments are conducted with support vector machines designed with sequential minimal optimization (SMO-SVM) function. Performances are analyzed individually and in combinations. Obtained results have exhibited the existence of dialect information at excitation source information and complementary cues at vocal tract system.