Combination of temporal trajectory filtering and projection measure for robust speaker identification

This paper presents a method that combines the techniques of temporal trajectory filtering and projection measure for robust speaker identification. The proposed robust feature, called Relative Autocorrelation Sequence Mel-scale Frequency Cepstral Coefficients (RAS-MFCC), is derived based on filtering the temporal trajectories of short-time one-sided autocorrelation sequences. This filtering process can minimize the effect of additive noise in the noisy speech. Since the norm of RAS-MFCC shrinks due to noise corruption, the projection measure (PM) technique, which is effective in dealing with the norm shrinkage of cepstrum, can be applied for the distance measure of RAS-MFCCs. The combination of these two techniques is then applied to a task of speaker identification of 100 speakers. Our experiment shows that the use of RASMFCC feature achieves significant improvement in identification rate as comparing with the use of MFCC. The combination of RAS-MFCC feature with PM technique can further improve the recognition accuracy.