Automatic language identification and discrimination using the modified group delay feature

Automatic language identification (LID) systems use features derived from the Fourier transform magnitude like MFCC, its derivatives and also PLP cepstra. Though half of the underlying spectral information is discarded in these cases, attempts to utilize the phase spectrum for deriving features have been minimal. This paper investigates the use of features derived from the Fourier transform phase for implementing LID systems. Features derived from the modified group delay function which we call the modified group delay feature (MODGDF) are used in this study. Performance of the MODCDF and the traditional MFCC for a GMM based LID system for a 3 and 1.1 language task are discussed. Results of language discriminability analysis are also presented. The MODGDF is found to outperform MFCC in terms of both performance and discriminability of languages.