Fast Independent Component Analysis (FastICA) in Separating Vocals and Instruments in the Art of Geguntangan

Jurnal Elektronik Ilmu Komputer Udayana Volume 8, No 3. February 2020 p-ISSN: 2301-5373 e-ISSN: 2654-5101 Fast Independent Component Analysis (FastICA) in Separating Vocals and Instruments in the Art of Geguntangan a1 Angga Pramana Putra , I Gede Arta Wibawa a2 a Informatics Department, Faculty of Mathematics and Natural Sciences, Udayana University, Jalan Raya Kampus Unud, Jimbaran, Bali, 80361, Indonesia 1 Abstract Geguntangan is pesantian in religious ceremonies in Bali accompanied by gamelan music. The human sense of hearing tends to have limitations, which causes not all vocals mixed with gamelan to be heard clearly. Therefore we need a system that can be used to separate vocals with gamelan in the geguntangan. Separation of sound sources is categorized as Blind Source Separation (BSS) or also called Blind Signal Separation, which means an unknown source. The algorithm used to handle BSS is the Fast Independent Component Analysis (FastICA) algorithm with a focus on separating the sound signal in a wav-format sound file. FastICA algorithm is used for the sound separation process with the value parameter used is Mean Square Error (MSE). From the simulation results show the results of MSE calculations using the mixing matrix [0.3816, 0.8678], [0.8534, -0.5853] obtained the -5 -6 results for the FastICA method, the MSE value is 3.60 x 10 for the vocal and 1.71 x 10 for the instrument. Keywords: Blind Sources Separation, Fast Independent Component Analysis, Audio Signal Processing, Mean Square Error 1. Introduction Geguntangan is pesantian in religious ceremonies in Bali accompanied by gamelan. Geguntangan often used in religious traditions to accompany the ceremony and also for public entertainment. However, the voice mixed between vocal and gamelan causes difficulty in learning the vocal in the geguntangan. The human sense of hearing tends to have limitations, which causes not all vocals mixed with gamelan to be heard clearly. Therefore we need a system that can be used to separate vocals with gamelan in the geguntangan. Separation of sound sources is categorized as Blind Source Separation (BSS) or also called Blind Signal Separation, which means an unknown source [1]. BSS is a way to separate the mixed signals into several forming signals, without information about the number of signal sources, or the process of mixing the signals. BSS utilizes the different signal characteristics before the sensor is detected and the information obtained due to differences in the angle of arrival and distance of the sensor. The sensor used is a microphone. There are two techniques in recording sound, namely, single-channel and multichannel. Single-channel is a recording technique that uses a single sensor, and multichannel is a recording technique that uses more than one sensor. The algorithm used to handle BSS is the Fast Independent Component Analysis (FastICA) algorithm. This algorithm focuses on the separation of sound signals in a Wav format sound file that has two or more sounds mixed so that separate sound results are recorded. The reason for using the wav format is that the wav format file contains sounds that are not compressed [2]. This research uses the FastICA Algorithm. FastICA algorithm can distinguish the elements or components of the signal mixture independently [3], which in this research uses only two sound sources, namely vocal sound signal and instrument. Before doing the FastICA process, the first thing to do is to do some pre-processing. Pre-processing in question is centering and whitening. 219 Putra & Wibawa Fast Independent Component Analysis (FastICA) in Separating Vocals and Instruments in the Art of Geguntangan Besides that, there is a Non-negative Matrix Factorization (NMF) algorithm, which can also be used in Blind Source Separation (BSS). NMF can process data on a large scale using a matrix factorization model compared to the classical algorithm. Some of the advantages of NMF are easy implementation, good interpretability of decomposition results, and small storage space [4]. Most of the research on the NMF method to overcome the limitations of the NMF method, namely the fact that NMF is a temporal magnitude model only [5]. Another algorithm commonly used to handle the Blind Source Separation (BSS) approach for signal processing problems is the Sparse Component Analysis (SCA) algorithm. SCA utilizes less signal to extract the source and has higher precision in terms of signal separation. If the signal source has a Gaussian distribution, which is a noise model that follows a standard normal distribution with zero averages, the SCA can still extract rare sources effectively. SCA is also a promising approach for BSS when there are fewer sensors from the source [6]. But, FastICA can sparate more effectively than SCA because the FastICA algorithm can define how many and what specifically source will be separate by using mixing matrix. The FastICA algorithm is used for the sound separation process with the value parameter used is Mean Square Error (MSE) or see the similarity between the output results with the input selected to test the sound output results. In speech signal recognition, the process of verbally listening to the output results compared to voice input is used to measure the parameters of success. Besides that, by comparing the input and output signals, it can be seen the results of the sound separation process. 2. Related Work Blind Source Separation is one way to blindly separate a mixed-signal into several forming signals [7]. BSS is one of the techniques used to obtain sources from blind mixing. Because every mixed component can be reconstructed again into its constituent signals, many algorithms can be used to solve problems in Blind Source Separation. FastICA algorithm in its testing of Blind Source Separation is better than the ability of the PCA algorithm, and NMF [8]. In testing the algorithm to determine a superior algorithm between FastICA, PCA, and NMF using the parameters of signal to interference ratio (SIR), signal to distortion ratio (SDR) and signal to artifact ratio (SAR). The greater the value of the parameter, the better the algorithm used for the BSS method, and vice versa. By using the three parameters, the FastICA method is superior to the other two methods tested. There are two types of sound sources used in the FastICA Algorithm, namely single-channel and multichannel. FastICA can run using both kinds of sound sources by modifying it with FSS-Kernel (Finite Support Samples Kernel), where nonlinear functions are replaced with PDF (Probability Density Function). The results of the study are that FastICA with FSS-Kernel modification can effectively separate sound sources with more than one mixed-signal [9]. FastICA can be used to improve sound quality by reducing noise or noise in an audio signal. This can be done by separating noise with sound signals. Another method that is usually used to improve the qu (...truncated)