A Machine Learning Approach to Automating Bengali Voice Based Gender Classification

Abstract

Does thicker vocal folds produce sounds with longer wavelength? And can they produce higher pitches to human ears? We address these types of questions and try to identify the difference between male and female voices. By using Machine learning algorithm it's possible to identify the gender from voices. And for that we extract voice signal's MFCCs features by calculating Discrete Fourier Transform, Mel-spaced filter-bank and log filter-bank energies. Identify gender from natural voice can be one of the most important part of voice recognition. In normal voice to text conversion it's not important to detect the voices gender. But when we use this voice recognition for real life applications, it will be densely needed to identify the voices gender. Gender identifying from voice is a field of Natural language processing which is a branch of artificial intelligence. We followed a simple working sequence for getting the ultimate result. The sequence is, Input-audio-file, Pre-works, Feature Extraction, Creating CSV file with features, Train the model and finally test with test data. For feature extraction we used Mel-frequency cepstral coefficient (MFCC). And for mapping and selection we used Logistic Regression, Random Forest and Gradient Boosting. After all this work we get 99.13% accuracy on the dataset that containing 1652 data of more than 250 speakers and tested them with 400 male and 400 female voices.