<![CDATA[Human voices vary from person to person. Men usually have larger vocal folds than women so their voice tend to be lower. Today virtual assistant and voice-based chatbot are still unable to differentiate gender based on human voice whereas if the user's gender could be known we can use it to understand behaviours of a particular gender. Learning Vector Quantization (LVQ) version 1 is used in this research as a method to classify human voices with two classes which are male and female. Sound characteristics that used as features in this research are energy, zero crossing rate, entropy of energy, spectral centroid, spectral spread, spectral entropy, spectral flux, and spectral rolloff. Highest result are at 75,5% when using 10 as maximum epoch, 0.1 as learning rate, and Normalized Cross Correlation as similarity measurement. Accuracy when using Normalized Cross Correlation to measure similarity is at 75,5% thus making it higher compared to Euclidean distance and Manhattan distance which only get 74,4% accuracy both. This research also tested using K-fold Cross Validation with 5 folds and highest accuracy obtained when testing fourth fold at 75,6%. Therefore, this research also used Recursive Feature Elimination to determine impacts of sound features on accuracy resulting best feature is spectral entropy whilst worst features are zero crossing rate, spectral rolloff, and spectral centroid.]]>
Copyrights © 2020
