One of the best forms of expression for emotion is music. The popularity of music streaming services and the usage of automated song recommenders have sparked a lot of curiosity about the power of music to evoke strong emotions. Several ML algorithms and acoustic and lyric elements taken from music data samples are combined in Music Emotion Recognition (MER) methodologies. However, the fact that no one emotion is depicted in each piece of music is the largest obstacle to accurately forecasting musical emotions. This paper explores the integration of acoustic and lyric features to enhance emotion prediction in music using machine-learning approaches based on the PMemo dataset. The methodology involved extensive preprocessing, including text cleaning, tokenisation, and Word2Vec embedding for lyrics, along with data balancing using SMOTE and feature scaling. Machine learning models such as XGBoost and AdaBoost were used for classification tasks on valence and arousal. Results revealed high training accuracy, particularly for audio-based classification (XGBoost achieving 99.72% for valence), but moderate testing performance (63.54% for arousal). The lyrics-based emotion prediction achieved 92.95% training accuracy but 56.57% testing accuracy. The proposed solution demonstrates significant improvements in arousal and valence detection across multimodal datasets, outperforming existing models in classification accuracy.