Optimizing the configuration of deep learning models for music genre classification

Music genre categorization is a fundamental use of sound processing methods in the realm of music retrieval. Typically, people are responsible for categorizing music genres. Machine learning approaches can automate this procedure. Therefore, in recent years, several approaches have been suggested to achieve this objective. Nevertheless, the given findings indicate that there is still a discrepancy between the observed results and an optimal categorization method. Hence, this paper introduces a novel approach for accurately forecasting music genres by using deep learning methodologies. The proposed approach involves preprocessing the input signals and then representing the characteristics of each signal using a combination of Mel Frequency Cepstral Coefficients (MFCC) and Short-Time Fourier Transform (STFT) features. Subsequently, a convolutional neural network (CNN) is applied to process each group of these characteristics. The proposed technique utilizes two CNN models to analyze MFCC and STFT data. Although the structure of these models is identical, the hyper-parameters of each model are individually adjusted using the black hole optimization (BHO) algorithm. Here, the optimization method finetunes the hyperparameters of each CNN model to minimize their training error. Ultimately, the results of two Convolutional Neural Network (CNN) models are combined to determine the music genre using a classifier based on SoftMax. The efficacy of the suggested methodology in categorizing music genres has been assessed using the GTZAN and Extended-Ballroom datasets. The experimental findings demonstrated that the suggested approach achieved classification accuracies of 95.2 % and 95.7 % in the two datasets, respectively, indicating its superiority over earlier efforts.