Identity Feature Extraction Scheme of Curvelets for Speaker Recognition

This paper produces two types of the features of speaker recognition, mean of column elements (MC) and squared 2-norm of column elements (SNC). Both of them are derived from the curvelets representing the geometrical structure of squared modulus of Gabor representation of one-dimensional speech. The performance evaluation experiments have been conducted and the results indicate that with the score of 87.23%, the feature of mean of column elements bears a little better identification effect than that of squared 2-norm of column elements. The recognition accuracy of Mel-frequency cepstral coefficients (MFCC) reaches 86.52% based on the same speech database