Modeling very long-range structures and complex interdependencies in musical pieces remains challenging. This paper proposes an enhanced model for symbolic music generation, specifically focusing on piano compositions, by integrating a self-attention mechanism with an LSTM network. The self-attention layer allows the model to weigh the importance of past hidden states, thereby potentially capturing more nuanced long-range dependencies and contextual information within the music. We utilize the extensive Lakh MIDI dataset (LMD), focusing on its piano components and employ a piano-roll representation for training. The proposed LSTM-Attention model, within an adversarial framework detailed subsequently, is designed to generate coherent and expressive piano melodies and harmonies. This work demonstrates the potential of self-attention to enhance the quality and structural coherence of music generated by LSTM-based systems.