Generating Emotional Music Based on Improved C-RNN-GAN

This study introduces an emotion-based music generation model built upon the foundation of C-RNN-GAN, incorporating conditional GAN, and utilizing emotion labels to create diverse emotional music. Two evaluation methods were employed in this study to assess the quality and emotional expression of the generated music. Objective evaluation utilized metric calculations, comparing the generated music to the music in the EMOPIA database, including factors like note range, chord count, and chord consistency. Additionally, subjective assessment involved inviting 20 listeners to hear a set of both real and generated music. Listeners were asked to distinguish between real and generated music and evaluate emotional expression and harmony. The results indicate that the model-generated music successfully conveys a variety of emotions and approaches the quality of human-composed music.