MUSIC GENRE CLASSIFICATION USING NEURAL NETWORKS

Meet Raval

doi:10.26483/ijarcs.v12i5.6771

Abstract

In recent years, the complexity of making music has lessened, resulting in many individuals making music and submitting it to streaming media. Because of the huge music streaming media, people are spending a lot of time seeking for certain songs. As a result, the capacity to swiftly categorise music genres has become increasingly important. As machine learning and deep learning technologies progress, convolutional neural networks (CNN) are being employed in several fields, and several CNN-based versions have emerged one after the other. Traditional music genre classification necessitates professional abilities to manually extract features from time series data. We developed a music genre categorization model using CNN's audio advantages and features to save users time while searching for different types of music. During the pre-processing, Librosa is used to convert the original audio files into Mel spectrums. The Mel spectrum is transformed and supplied into the suggested CNN model for training. On the GTZAN dataset, the 10 classifiers' decisions are subjected to a majority vote, with an average accuracy of 84 percent. Music genre categorization using neural networks (NNs) has seen some modest success in recent years. The success of song libraries, machine learning techniques, input formats, and the types of NNs utilised has all been mixed. This article looks at some of the machine learning approaches utilised in this sector. It also involves research on musical genre classification. Images of spectrograms produced from time slices of songs are fed into a neural network (NN) to classify the songs into different musical genres.

Highlights

Since the Internet's birth, many people have uploaded music previously recorded on vinyl records or CDs to streaming media on the Internet
Many of the papers that employed convolutional neural networks (CNN) compared their models against other machine learning techniques including k-neural networks (NNs), Gaussian mixtures, and SVMs, and found that CNNs beatthem all.Our research revealed that when considering the whole 30s track duration, current state-of-the-art algorithms perform with an accuracy of around 91 percent, well above human skills for genre identification
Many of the papers that employed CNNs compared their models against other machine learning techniques including k-NN, Gaussian mixtures, and SVMs, and found that CNNs beat them all.We found some representation learning research in other applications, such as chord recognition and music starting detection, that is still within the scope of content-based music informatics

Summary

Introduction

Since the Internet's birth, many people have uploaded music previously recorded on vinyl records or CDs to streaming media on the Internet. People look for popular music through music streaming services, which have become increasingly popular in recent years, and the enormous internet music library makes finding certain genres or songs difficult. Because most music in today's music streaming media just has a title and a creator, and most of it isn't tagged. This makes it harder to find hidden tags in songs and categorise them into genres. Machine learning has gained a lot of popularity in recent years. Depending on the type of application and the data set available, different types of machine learning algorithms are better suited for different applications

Objectives

Methods

Findings

Conclusion