Jiaying Li, Nat Condit-Schultz
I attended the poster session at SMPC2022, and my research was focused on chord progression similarity in Chinese Popular Music.
With the development of the Internet and media platforms, people have access to more and more music. The criticisms of “reusing” chord progressions in Chinese popular music are intensifying in recent years. During my research, we investigated the reuse of identical and similar chord progressions in Chinese popular music in the past ten years. We purposed a novel Chord Progression Similarity Index (CPSI) based on the Markov Model of pitch content in consecutive chords.
To begin with, we sampled a Chinese popular music database, which contains the 200 most popular songs in the past ten years. We accessed and checked the existing chord transcriptions of 200 sample songs. After encoding the chords from the verse and choruses of each track, we transposed all major pieces to C major and minor pieces to A minor.
The CPSI is calculated from the Euclidean distance of the transition matrix formed by the notes required for the chord formation in the chord progression. Compared with the chord alignment method, the advantage of using the transition matrix to calculate the similarity index is that this algorithm can handle chord progressions containing different chord numbers and is not influenced by inversion chords. By applying the algorithm to the Chinese popular music dataset we sampled, the result shows that the CPSI in Chinese music keeps decreasing, which means that more and more similar chord progressions are used, especially in the recent three years.
Based on the initial ideas mentioned above, three different transition matrix algorithms are derived, and their differences are mainly in the calculation of the chord weights in the chord progression: 1) Leading chords have higher weight; 2) The transition between two similar chord weights lower; 3) The repeated chord progressions with slightly changes have lower weights.
To figure out how well the similarity index calculated by the algorithms matches the similarity perceived by the human ear, we also conducted a perceptual experiment. An additional adjacent matrix will be added to match the human perception to improve our model.