Automatic Speaker Recognition Using Multi-Directional Local Features (MDLF)

Abstract

A new feature called multi-directional local feature (MDLF) is proposed and applied in automatic speaker recognition. In order to extract the MDLF, a windowed speech signal is processed by fast Fourier transform and passed through 24 mel-scaled Filter Bank, followed by log compression. A three-point linear regression is then applied in four different directions, which are horizontal (time axis), vertical (frequency axis), 45◦ (time–frequency) and 135◦ (time–frequency). MDLF holds the characteristics of the speaker in time spectrum and results in better performance. In the experiments conducted, a Gaussian mixture model (GMM) with a different number of mixtures is used as the classifier. Experimental results show that the proposed MDLF has better recognition accuracy than the traditional MFCC features. The MDLF achieves excellent results both in text-dependent and text-independent speaker recognition, and in Arabic and English speech. The proposed technique is also language independent.