Music Emotion Detection using Weighted of Audio and Lyric Features

Abstract

Music emotions can be seen from the audio and lyrics features. Audio is signal data while lyrics are text data. Combining these two features is needed for detecting music emotions. This research used synchronized dataset of chorus audio and lyrics. Audio features that extracted include dynamics, rhythm, timbre, pitch, and tonality features. While the lyric features that extracted are psycholinguistic, stylistic and statistical features. Audio and lyrics features have preprocessing, data normalization and categorization processes. The normalization process used Min-Max Normalization method and the categorization process uses a Rule Based method. Detection of musical emotions is done by weighting the audio and lyric features of the Naive Bayes probability value. From the weighting of these features, we known that audio feature is a dominant feature then a lyric feature. The weighting ratio is 80% for audio features and 20% for lyric features. The accuracy of system using weighting is 0.774. It increased from the accuracy of system without any weighting.