Single Speaker Recognition Using Deep Belief Network Gender Classification Voices
Murman Dwi Prasetio, Shinya Sekizaki, Tomohiro Hayashida, Agus Susanto, Ichiro Nishisaki, Muhammad Abdillah · 2019
Recently, the algorithm of machine learning are used to able to train, enhance, characterize and anticipate the data results accurately. In a way to training, the process on the algorithm can be able to produce an appropriate model based on that data; it's like supervised and unsupervised data. In this paper, we tried to trace the gender (male and female) from acoustic data, i.e., pitch, median, frequency etc. The gender that would like to implement is classified on the basis of the intensity of their utterances. To analyze the utterances, the voice intensity measuring by the hamming window to make a normalize curve to obtain the peaks of the utterances where peaks are found from each frame of speech utterance when it is divided into frames of the length of 20 milliseconds. At certain amplitude levels it can be considered to find a peak. As well as making decisions about gender use a thresholds that are adapted are adjusted. If the area of an utterance is above the threshold the gender type is a female otherwise male. After that, we handle the feature learning from the utterance into deep belief network as a machine learning tool to predict single speech by gender classification voices with optimization (taboo search) to train several neurons in the initial weight vector for the accuracy of female and male voices 75.67% and 80.83% precisely.