Text Independent Speaker Identification Model Using Mixture of Pearsonian Type I Model and K-Means Algorithm

Title: Text Independent Speaker Identification Model Using Mixture of Pearsonian Type I Model and K-Means Algorithm
Publisher: Guru Nanak Publications
ISSN: 2278-0947
Series: Volume 3 Issue 2
Authors: K.Srinivasa Rao , P.Chandra Sekhar and M.Sesha sayee


Abstract

Speaker identification model plays an important role in authentication and recognition of persons. Much work has been reported in literature regarding speaker identification models using Gaussian mixture model. The Gaussian mixture model has a prime drawback since it can characterize only mesokurtic and symmetrically distributed features. But in many practical situations the feature vector associated with each speech spectra (Mel-cepstral coefficients) may not be having a mixture of Gaussian components. Hence to have accurate the identification and recognition system, it is needed to develop generalize text independent speaker identification system . Hence in this paper we develop and analyses speaker identification system using mixture of Pearsonian Type I system of models. The model parameters are derived in updated equations EM algorithm. The initialization of model parameters of EM algorithm is obtained by K-means algorithm and moment estimates. Using likelihood function under the baysian frame work, the speaker identification algorithm is developed. The performance of algorithm is studied using ROC curves, confusion matrix and other quality matrices. It is observed that the system outperforms the model based speaker identification system models.

Keywords

Bayesian frame, EM algorithm, Gaussian mixture model, Likeliood function, Meso kurtic, Pearsonian type – I system, ROC curves, Speech spectra, Speaker identification model.

Download Full Text

(For complimentary copy, please contact Chief_editor@innovationjournals.com)