Recognising Bonang Barung Gamelan Instrument Playing Technique Using Convolutional Neural Networks

Abstract

The diversity of Indonesian culture is an interesting thing to be reviewed further. One of them called karawitan, which uses gamelan instruments as its medium with various playing techniques on each instrument. This can be found on the bonang barung instrument which has at least thirteen different playing techniques or known as tabuhan techniques. However, it is not easy for beginners to learn karawitan, since there are many techniques that must be learned on each instrument. Besides wanting to help beginners learn karawitan with the support of up-to-date system capabilities in processing data, this research is also expected to enrich research in the field of audio classification. Types of feature extractions such as mel spectrogram and MFCC were tested on the CNN architecture. In addition, the process of cleaning noise and levelling the loudness level of the raw data is applied with the aim of getting better audio quality. Apparently, at the best hyperparameter settings, it was found that the MFCC feature is better for audio data containing noise, which achieves an accuracy as good as 99%, while Mel Spectrogram excels at noise-free audio data with an accuracy as good as 98%. Therefore, the end of this study shows that the MFCC and Mel Spectrogram features have their respective advantages.