A Wavelet Packet and Mel-Frequency Cepstral Coefficients-Based Feature Extraction Method for Speaker Identification

One of the most widely used approaches for feature extraction in speaker recognition is the filter bank-based Mel Frequency Cepstral Coefficients (MFCC) approach. The main goal of feature extraction in this context is to extract features from raw speech that captures the unique characteristics of a particular individual. During the feature extraction process, the discrete Fourier transform (DFT) is typically employed to compute the spectrum of the speech waveform. However, over the past few years, the discrete wavelet transform (DWT) has gained remarkable attention, and has been favored over the DFT in a wide variety of applications. The wavelet packet transform (WPT) is an extension of the DWT that adds more flexibility to the decomposition process. This work is a study of the impact on performance, with respect to accuracy and efficiency, when the WPT is used as a substitute for the DFT in the MFCC method. The novelty of our approach lies in its concentration on the wavelet and the decomposition level as the parameters influencing the performance. We compare the performance of the DFT with the WPT, as well as with our previous work using the DWT. It is shown that the WPT results in significantly lower order for the Gaussian Mixture Model (GMM) used to model speech, and marginal improvement in accuracy with respect to the DFT. WPT mirrors DWT in terms of the order of GMM and can perform as well as the DWT under certain conditions.