Article Preview
TopIntroduction
Music features are characteristic classifications that are used to distinguish between different genres of music. Also, each genre differs from other genres in certain musical features. In the pre-computational intelligence era, music feature categorization has traditionally been performed manually, mostly due to the lack of modern human-computer interaction concept, and obviously for the lack of enough computational processing abilities of the computers. However, with the ever-increasing number of digital music and vast features, feature recognition using the neural the network is producing a wide range of results across a variety of experiments recently. Nowadays, music feature classification is a popular topic for research, particularly in the fields of Music Information Retrieval (MIR) and Neural Network. Since Artificial Neural Network with a variety of learning methods have been introduced in this domain, there have been many research works done on methods for classifying music features according to the diversity of their features. The human brain computes in a distinctly different way rather than conventional computers. Keeping this concept in mind, the research work on Artificial Neural Network has been motivated. The main ideology of a Neural Network is to represent a linear or nonlinear and parallel computing architecture. Researchers have also suggested methods for music feature classification using multiple feature vectors and pattern recognition ensemble approach. In figure 1, the authors represent the process blocks of traditional music classification techniques utilizing neural networks. Music segments are also decomposed according to time segments obtained from the beginning, middle and end parts of the original music signal (time-decomposition), alongside with distinguished musical features such as pitch, intensity, beat, drop and so on, which is unlikely to be the same for even two distinct music files. In the present scenario, one of the biggest bottlenecks in many Music Information Retrieval (MIR) tasks is the access to large amounts of music data and their features, in particular to audio features extracted from commercial music recordings (Porter et al., 2015). However as of now, at least hypothetically these challenges can also be addressed by using a multilayer neural network with more than one hidden layer within it, and with backpropagation feedback signaling. Previous work in this topic represented a robust hypothesis to show how the music genre recognition can be done from distinct musical features using a single-layered feedforward neural network (Das & Kolya, 2019). In this paper, the authors carry forward the problem domain further, to propose an empirical technique of successfully tackling the feature diversity of music. They present a simple yet featureful work to show the performance of a shallow neural
network or single-layered neural network on an already established music feature corpus with unsupervised learning. They focus on the empirical computation that takes place while the network learns basic but distinct music properties such as beat, Fourier signal transmission, Mel-frequency signal, and finally pitch of the tunes. These music feature recognitions can be done from features converted as distinct binary inputs with the Hebbian learning technique, which is at the core of the proposed approach. The authors are here interested in observing and presenting the intricate differences like such signals when they pass through a single layer network following and unsupervised learning pattern. They further examine and observe each music features passing to their respective perceptrons and behavioral expression through the activation summarizations. Finally, they visually observe and note the performance of the proposed method to classify the errors. The authors achieve an overall accuracy of 90.36% in the validation phase. From that, they also perform a comparative evaluation with such state-of-the-art works in a similar domain and propose the paths for future enhancements.
Figure 1. Ground-level music classification using neural networks
Main contributions of this paper are: