Enhancing Music Emotion Recognition with LSTM: Evaluating Various Embedding Techniques
Affreen Ara, Rekha V · International Journal of Electronics and Communication Engineering · 2025
The study investigates the application of Long Short-Term Memory (LSTM) networks for emotion classification in music lyrics. It focuses on the comparative effectiveness of various word embedding techniques. It evaluates the performance of static embeddings (GloVe, Word2Vec, FastText) versus contextual embeddings (BERT, Distil BERT) across three datasets: MER Lyrics, Mood Lyrics, and Combined Lyrics. Additionally, the study examines the role of stylistic and content-based features in enhancing classification accuracy. The results demonstrate that contextual embeddings considerably outperform static embeddings, achieving accuracy rates of up to 98% compared to 60% for static approaches. Moreover, combining multiple lyric datasets leads to improved model generalization. The findings show the potential of transformer-based models for advancing music emotion recognition. Future research will focus on optimizing large embedding models using techniques such as pruning, quantization, and distillation to enhance computational efficiency.