Does Harmony Influence Musical Style/Genre?

Camille Noufi and Brandon Jacquez

Harmonic analysis of popular music allows us to unveil patterns among artists that continue across decades. Theorist Walter Everett used a six category system to describe these patterns for a different pop/rock dataset. A link to a table summary of his findings can be found here. This work can be tedious and subjective, but the McGill Billboard dataset aimed to provide accurate data to analyze the top songs on the Billboard’s list. We, a group of students in CU’s Computation Musical Analysis class, used the McGill dataset to develop our own findings on this topic to compare to Everett’s. We also compared our cluster analysis results with categorical results from a previous study (slides 10–17) on McGill done by the class two years ago.

With this knowledge in mind, we wanted to re-evaluate these categories using the McGill Billboard dataset and K Means clustering analysis. We began by repurposing work of a previous group of students in our class who looked at the transitional probability of one chord to another in the McGill dataset. We used the transitional probability for each song as our data for a K Means cluster analysis using 6 means with 1,000 iterations. Our code was written in python and the clustering was done using the sciKitLearn framework. We found many similarities between all three studies. We used information from a study done by Ashley Burguoyne (a creator of the McGill dataset) to investigate the emergent clusters’ relations to release year. This helped us dig deeper into the results. Our findings indicate that relation of harmonic progressions to musical genre and style may not be as cut and dry as one would think.

Below is a graphical visualization of the transitional probabilities of a chord leading to another chord, labeled with our interpretation of the most dominant features of each cluster. These interpretations are largely found from schemas detailed at http://openmusictheory.com/popRockHarmony.html. To read the graph, compare the relative heights of the rectangles in a particular column. In a specific column, the rectangle order tells you which destination chord you are landing on from the column’s origin chord. The rectangles’ heights are “large” if the particular destination chord is more likely to be transitioned to than others from the column’s origin chord.

Cluster 0: Classical — Circle of Fifths

Insights:

• Primarily 60s and 70s. Like the doo-wop progression found in cluster 4, these earlier songs rely on simple classical progressions.

Cluster 1: I — IV — I — V — IV — I : Blues

Insights:

• The harmonic progression suggests a blues progression, but that does not necessarily mean the artist encompass the blues style or genre past the harmonic structure. The chords found in the blues progression (I, IV, V) are also the reverse circle of fifths when starting at I, as well as some of the simplest chords to play on a variety of instruments. Artist analysis shows a wide variety of musicians who most listeners would argue do not sing blues (ex. Cheap Trick’s “I Want You to Want Me”). This gives us an insight that harmonic structure can play a role in genre determination but other elements in a song play a more significant role in what the listener or artist believes the genre is.

Cluster 2: Random

Cluster 3: V/ bVII — IV — I : Plagal

Insights:

• This categorization is much more structural than interpretive of genre. A quick internet search on plagal progression does not show any significant ties this progression has to a genre or style. According to Everett, rather than as a function present throughout the whole song, a plagal progression seems to be used in the large “rock/pop” genre as chromatic harmony extending a passage.

Cluster 4: VI — IV — V — I : Classical — 50s Doo-Wop

Insights:

• The naming of a 50’s Doo-wop progression should be taken with a grain of salt. The 50’s may contain the most simple examples of this progression, but our data showed it most common of 70’s music. “Hocus Pocus” by Focus is a staple of 70’s music that doesn’t have that stereotypical 50’s vibe. Given the cluster 5 insights below, it might be more apt to classify this progression as being common of the 70’s. 50’s Doo-wop has a very simple structure that makes it easy to identify and play. By the 70’s, it’s still quite common but in a more creative way that isn’t as obvious.

Cluster 5: Predominant -> bVII/V -> I : Modal or Rock Tonal/Minor

Insights:

• The majority of the songs are in minor or flux between major and minor. Many of these artists fall into the rock genre of the 1980s and 1990s. Some 1970s music is present in this cluster and a few songs from the 1960s are as well. Diversity in chord use expanded from the 70s to the 90s; it makes sense that these songs using a variety of modes would fall in the later years of our data set.

Our graphs look similar to the classifications determined by Everett and the previous class. See Table below:

This dataset is meant to be a representative sample of American popular music. This already introduces some biases that need to be considered. When trying to find all the diversity and patterns in a set, you can’t reasonably do this analysis on every song ever written. But taking the most commercially successful might hinder diversity in the dataset; hiding potentially new or interesting ideas behind patterns that might be influenced by record companies with profit in mind.

Although no statistically significant conclusions can be made — our clustering analysis did not include confidence intervals and our data did not include “ground truth” genre — this study serves as a springboard for which future work could be done. Everett’s naming conventions were a standard for us to compare using computational music analysis tools but 6 clusters may not actually be the most ideal number. As detailed in a paired post, ““Harmonic Progression as Understood Through Cluster Analysis,” increasing the number of clusters might pull out more detailed conclusions.

Future studies could perform clustering on different elements in the dataset to determine any true correlation between harmonic progression and other factors that influence style, like year or metre. As stated in our paired post, harmonic progression does play a part in style (and perhaps, subjectively, genre) classification. However, several other elements not available in our dataset also define genre, such as timbre, lyrics, tempo, etc. Our classmates investigate these elements further with the Million Song Dataset, a more comprehensive dataset that includes non-harmonic parameters such as these for a huge amount of popular music. A future improvement to our study would be to combine our analyses for songs overlapping the two datasets by comparing more useful metadata than we had available with the essential harmonic data we developed.

If you’re interested in looking at the data yourself, you can clone our github repo here