Gamelan music onset detection using Elman Network

Gamelan, one of Indonesia´s traditional music instruments, generates signals that have variations in terms of fundamental frequency, amplitude, and signal envelope, due to its handmade construction and playing style. Therefore onset detection which is crucial for gamelan music analysis; undergoes several shortcomings using spectral and temporal features. This paper investigates the implementation of machine learning approach to understand statistical variations contained in gamelan signals which are relevant to onsets. The method uses Elman Network which consists of one hidden layer. Input units came from the power spectrogram and its positive first order difference of the signals as well as the context units from the output of each hidden unit one step back in time. The spectrogram was built using Short-time Fourier Transform and was converted into the log of Mel scale. A fixed threshold was used to select among the local peaks and the result is considered as binary classification of the signal at each time instant. The network was trained on a set of gamelan signals consists of synthetic and real recording data of single instrument playing. The performance gained 93% of F-measure.