Gender classification and speaker identification using machine learning algorithms

Abstract

The speech is a unique biological feature to each person, and this is commonly used in speaker identification tasks like home automation applications, transaction authentication, health, access control, among others. The purpose of the present work is to compare gender classification and speaker identification experiments in order to determine the machine learning algorithm that shows the best metrics performance based on Mel frequency cepstral coefficients (MFCC) as speech descriptive features. In this process, the machine learning algorithms implemented were logistic regression, random forest, k-nearest neighbors and neural network, which were evaluated with accuracy, specificity, sensitivity and area under the curve. The schemes that revealed the best performance were random forest and k-nearest neighbors, reflecting an AUC (area under the curve) of 1, which indicates that the models have robust capacity of classification both in isolated samples and in complete audio files. The results obtained open guidelines to carry out another type of experimentation using the MFCC features with audios where the environment noise factor is included to measure the performance with these classification algorithms. The experimentation proposed for this work seeks to be applied in the future in different areas, where MFCC are used to describe the voice to perform another type of classification.