Vision Transformer for Music Genre Classification using Mel-frequency Cepstrum Coefficient

The music industry has gone through some seismic changes since the dawn of the Internet. It has been reaching a much larger audience because the appetite for different music styles has increased. New songs are written and released every day, which makes genre classification a tiresome and lengthy task. A good music application should be able to recommend songs according to the user's preferred genre and identify the genre of the songs with precision. In a successful attempt towards the same, we leverage a transformer-based model to outperform the frequently used CNN model. In this research, acoustic features (MFCC) are extracted from the audio files from the FMA dataset for genre classification purposes. The songs represented by MFCC are trained and evaluated on the novel Vision Transformer, RNN-LSTM, and CNN-based architecture.

I. Introduction

The Internet has democratized the production and distribution of music worldwide and has led to an explosion in the amount of music released online. To process such large amounts of audio data, we need better Music Information Retrieval (MIR). MIR is a multidisciplinary field that combines digital audio signal processing, pattern recognition, software engineering, and machine learning. MIR helps build better recommendation systems and helps classify songs based on the genre, mood, etc. MIR uses many different kinds of features to effectively classify audio files, one of which is the Mel-frequency Cepstral Coefficient(MFCC), which we will use in our research to classify genres from the FMA [1] dataset.