Exploration of Music Emotion Recognition Based on MIDI

Audio and lyric features are commonly considered in the research of music emotion recognition, whereas MIDI features are rarely used. Some research revealed that among the features employed in music emotion recognition, lyric has the best performance on valence, MIDI takes the second place, and audio is the worst. However, lyric cannot be found in some music types, such as instrumental music. In this case, MIDI features can be considered as a choice for music emotion recognition on valence dimension. In this presented work, we systematically explored the effect and value of using MIDI features for music emotion recognition. Emotion recognition was treated as a regression problem in this paper. We also discussed the emotion regression performance of three aspects of music in terms of edited MIDI: chorus, melody, and accompaniment. We found that the MIDI features performed better than audio features on valence. And under the realistic conditions, converted MIDI performed better than edited MIDI on valence. We found that melody was more important to valence regression than accompaniment, which was in contrary to arousal. We also found that the chorus part of an edited MIDI might contain as sufficient information as the entire edited MIDI for valence regression.