New Features for Automatic Text Independent Language Identification

NEW FEATURES FOR AUTOMATIC TEXT INDEPENDENT LANGUAGE IDENTIFICATION A. Nagesh1 and V. Kamakshi Prasad2 1Mahatma Gandhi Institute of Technology, Hyderabad, India E-mail: [email protected] 2Jawaharlal Nehru Technological University, Hyderabad, India E-mail: [email protected] The objective of this paper is to explore new feature vectors for Automatic Text Independent Language Identification (LID) system. The task of the automatic language identification (LID) is to quickly and accurately identify the language being spoken. The basic requirement to build a LID system is feature vectors. Most of the existing LID systems are based on MFCC feature vectors. In this paper, an approach is proposed to derive a new type of feature vectors from speech signal for LID task. The variation in frequency of occurrence of phonemes is effectively captures by using Gaussians. The underlying probability function of Gaussian Mixture model (GMM) represents variations in the form of Gaussian probabilities. The new feature vectors are obtained by transforming 12 dimensional MFCC feature vectors to 15 dimensional feature vectors by considering the probabilities of Gaussian mixtures. This feature vectors are modeled using Hidden Markov model for LID. The experiments were carried out by varying the number of states and mixture components in HMM. The LID system is evaluated for 10 languages. The six languages identification performance is 100%. This approach outperformed the existing conventional MFCC feature vector based LID systems. We have demonstrated the automatic LID studies on OGI_MLT speech database.