A novel approach to language identification using modified polynomial networks

Abstract

In this paper, a new method of classifier design, viz., Modified Polynomial Networks (MPN) is developed for the Language Identification (LID) problem. The novelty of the proposed method consists of building up language models for each language using the normalized mean of the training feature vectors for all the speakers in particular language class with discriminatively trained polynomial network having based on Mean-Square Error (MSE) learning criterion. This averaging process in transformed feature domain (using polynomial basis) represents in some sense the common acoustical characteristics of a particular language. This approach of classifier design is also interpreted as designing a neural network by viewing it as a curve-fitting (approximation) problem in a high-dimensional space with the help of Radial-Basis Functions (RBF) (polynomials in the present problem). The experiments are shown for LID problem in four Indian languages, viz., Marathi, Hindi, Urdu and Oriya using Linear Prediction Coefficients (LPC), Linear Prediction Cepstral Coefficients (LPCC) and Mel Frequency Cepstral Coefficients (MFCC) as input spectral feature vectors to the second order modified polynomial networks. Confusion matrices are also shown for different languages. © 2008 Springer-Verlag Berlin Heidelberg.