MemoMusic 3.0: Considering Context at Music Recommendation and Combining Music Theory at Music Generation

MemoMusic 3.0 enhances personalized music recommendation by considering the music listening context, and improves music generation by introducing music theory. One observation is that the context of music listening would affect the emotional states of listeners, positively or negatively. The other is that better music can be generated by introducing some music theory knowledge. We propose a Transformer-based music generation framework, which is trained into three models for Classic, Pop, and Yanni music respectively. The dominant melody of a music with expected Valence and Arousal values is used as a sample sequence to the model, and its output is adjusted according to music theory. Experimental results demonstrate that MemoMusic 3.0 performs better at improving the emotional states of listeners and achieves better user satisfaction.

I. Introduction

While music is universally enjoyed by human, it is also widely recognized as an effective way to regulate the emotional states of listeners [1]. The regulation mainly depends on three aspects, namely, the music, the listener, and the listener's perception of the music. First, music emotion is mainly delivered by the three musical elements of melody, rhythm, and harmony. Second, the listener's emotional states are much determined by internal factors such as the physical and spiritual health states of the listener, and greatly influenced by external factors such as the task or activity the listener is involved in. Third, music perception is a very subjective issue, which varies greatly from person to person. One person's trash is another's treasure. The listeners have their own favorite music genres and styles. Moreover, they have unique life experiences and memories.