Attention-based CNN-BiGRU for Bengali Music Emotion Classification

Abstract

For Bengali music emotion classification, deep learning models, particularly CNN and RNN are frequently used. To extract meaningful knowledge, however, past studies' shortcomings of low accuracy and overfitting have to be addressed. We have proposed a model combining Conv1D, Bi-GRU and the Bahdanau attention mechanism for music emotion classification of our Bengali music dataset. The model integrates distinct MFCCs wav preprocessing methods with deep learning methods and attention-based methods. The attention mechanism has increased the accuracy of the proposed classification model. The music is finally classified into one of the four emotion classes: Angry, Happy, Relax, Sad. The proposed Conv1D+BiGRU+Attention model is validated as more effective and efficient at classifying emotions in the Bengali music dataset than baseline methods, according to comparisons with baseline models. For our Bengali music dataset, the performance of our proposed model is 95%.

Instrumental Music Emotion Recognition with MFCC and KNN Algorithm

“…Mel-Frequency Cepstral Coefficients (MFCC) is one of the most common feature vectors used in speech and music related pattern recognition applications. MFCC has revealed its outstanding performance in speech and music emotion recognition [9] and [10]. The way MFCC works is based on the difference in frequency that is captured by the human ear so that it can represent the characteristics of sound signals as humans represent them.…”