Classification of music emotions with pre-trained models

Music is the most emotional of all arts and is therefore the closest and dearest to people. Emotions are a central part of human life, influencing how we think, act, and make decisions. Many scientists have studied emotions, including Paul Ekman, who classified basic emotions into six kinds: happiness, sadness, fear, anger, surprise, and disgust. This classification helps in understanding how emotions influence our behavior and daily interactions. Emotion analysis can benefit various fields such as healthcare, education, business, marketing, technology, and more. Various applications facilitate emotion extraction and analysis, including: • Text Analysis: Natural Language Processing (NLP) analyzes written texts to extract emotions. • Voice Analysis: Applications like Beyond Verbal and Affective analyze voice tone to extract emotions, which is useful in customer service centers. • Image and Video Analysis: Computer Vision techniques analyze facial expressions to identify emotions from images and videos. Music occupies a special place in stirring up and expressing human feelings, with its melodies, rhythms, and tones provoking a wide range of emotions. The analysis of musical emotions is a growing research area where scholars seek to understand how different musical elements influence listeners' feelings. My study focuses on classifying music sentiments using Deep Learning. To implement my proposed idea, I used some pre-trained models such as ResNet50, DenseNet121, and Inception V3 on two audio datasets after converting them to spectrogram images. The audio dataset was divided into ten-second segments, and two types of spectrogram transformations were used: classic spectrogram and Mel spectrogram. The outcomes demonstrated that the ResNet50 model achieved the highest performance on both datasets, although DenseNet121 had a higher accuracy with the Mel spectrogram using one of the data balancing techniques on the second dataset, but the performance of the ResNet50 model is still higher in terms of inference and generalization task. It also achieved the highest accuracy on the first dataset. This study provides insight into how Deep Learning techniques can be applied to analyzing and classifying musical emotions, contributing to a better understanding of music and emotions.