A Comparison Between MFCC and MSE Features for Text-Independent Speaker Recognition Using Machine Learning Algorithms

Abstract

Speaker identification is the process through which a person can be identified by using the physical characteristics of his voice. In recent years, this identification is done by using artificial intelligence algorithms along to feature extract methods from the speech signal such as the MFCCs (Mel Frequency Cepstral Coefficients). Text-independent speaker recognition consists of identifying a person by training and testing of the model with voice signals where one speaker will say different phrases. In the present research two types of audio features were extracted, on one hand the MFCC and on the other hand the MSE (Multiband Spectral Entropy). For the classification stage we use machine learning algorithms, such as k Nearest Neighbors, Random Forest, Deep Neural Networks, and Decision Trees. Two important databases of the literature were used in our experiments, LibriSpeech and ELSDSR. Four different experiments were defined: speaker identification in a group of 20 participants, speaker identification between men, speaker identification between women, and speaker identification by gender. Significant results were obtained when using the ELSDSR database, obtaining 93.99% precision in the experiment for speaker recognition by gender.