Speaker identification with vector quantization and k-harmonic means

A new method is proposed in this study to identify speakers in a relatively short time without decreasing success ratio. The method first extracts MFCC (Mel Frequency Cepstrum Coefficients) features. Then k-harmonic means is used to cluster samples before classification is performed by the nearest neighbor method. International HYKE database is used to test the performance of the proposed method in terms of success ratio and runtime, and compare with both MFCC+k-means and MFCC+LBG methods. Preliminary results show that the proposed method usually gives the best performance.