Music emotion recognition plays an important role in affective computing by enabling intelligent systems to interpret emotional patterns embedded within acoustic signals. This study proposes an attention-guided ensemble deep-learning framework termed ELCDA (Ensembled LSTM–CNN–DNN with Attention) for binary music emotion recognition using complementary acoustic feature representations. The framework integrates convolutional spectral learning, recurrent temporal dependency modeling, and dense statistical feature abstraction within a unified attention-based fusion architecture to capture rhythmic, harmonic, temporal, and spectral emotional characteristics simultaneously. The experimental evaluation was conducted on the publicly available Kaggle “Musical Emotions Classification” dataset containing 2,126 audio samples categorized into Happy and Sad emotional classes. After preprocessing, acoustic representations including MFCC, Chroma, waveform, Zero Crossing Rate, and Mel-Spectrogram features were extracted and standardized before model training. The dataset was evaluated using a 70:15:15 train–validation–test split along with repeated cross-validation experiments for stability analysis. Experimental results demonstrate that the proposed ELCDA framework achieved an accuracy of 85.5%, F1-score of 87.2%, and ROC-AUC of 0.91, outperforming conventional CNN, LSTM, and DNN baseline models in balanced emotional classification performance. Additional ablation analysis, sensitivity evaluation, SHAP interpretation, attention-weight analysis, and statistical validation tests further confirmed the robustness and effectiveness of the proposed framework. The study highlights the importance of multi-feature acoustic fusion and attention-guided representation learning for robust and interpretable music emotion recognition.
Zhang et al. (Mon,) studied this question.