NEURAL NETWORK BACKPROPAGATION FOR KENDANG TUNGGAL TONE CLASSIFICATION

Abstract

Kendang Bali is one of the instruments incorporated in this karawitan art. Balinese kendang can be played alone, called a kendang tunggal, where this type of game has a high level of difficulty understanding the tone of the Balinese drums played because some variations of the tone have similar sounds to other tones. Knowing the tone that is in the kendang song automatically can make it easier to learn it. The first approach method used to classify the tone of a kendang tunggal song is segmentation. The onset detection method is used to segment a kendang song with a variation of the hop size parameter. The segmented tone of the punch will be classified using the Backpropagation method. Feature values of autocorrelation, ZCR, STE, RMSE, Spectral Contrast, MFCC, and Mel spectrogram will be used in the classification process. This study performed variations in hop size values in onset detection and obtained the proper configuration at a value of 110. The addition of the normalization process to the onset detection method also helps the segmentation process of kendang songs correctly. The optimal backpropagation architecture obtained is learning rate 0.9, neuron hidden layer 10, and epoch 2000 produces an accuracy of 60.92%.