The need of advance Music Information Retrieval increases as well as a huge amount of digital music files distribution on the internet. Musical genres are the main top-level descriptors used to organize digital music files. Most of work in labeling genre done manually. Thus, an automatic way for labeling a genre to digital music files is needed. The most standard approach to do automatic musical genre classification is feature extraction followed by supervised machine-learning. This research aims to find the best combination of audio features using several kernels of non-linear Support Vector Machines (SVM). The 31 different combinations of proposed audio features are dissimilar compared in any other related research. Furthermore, among the proposed audio features, Linear Predictive Coefficients (LPC) has not been used in another works related to musical genre classification. LPC was originally used for speech coding. An experimentation in classifying digital music file into a genre is carried out. The experiments are done by extracting feature sets related to timbre, rhythm, tonality and LPC from music files. All possible combination of the extracted features are classified using three different kernel of SVM classifier that are Radial Basis Function (RBF), polynomial and sigmoid. The result shows that the most appropriate kernel for automatic musical genre classification is polynomial kernel and the best combination of audio features is the combination of musical surface, Mel-Frequency Cepstrum Coefficients (MFCC), tonality and LPC. It achieves 76.6 % in classification accuracy.