The Narcissism of Minor Differences

My data visualization project represents traditional Irish tunes visually to highlight their similarities. The idea for the project came from something my mother says to me regularly when I’m home visiting and playing the fiddle: “How do you keep all those tunes straight? They all sound the same to me”.

There are many sources of tune data online. Originally I’d planned to scrape some from thesession.org, but my friend, Pete, actually had a database of tunes he’d scraped several years ago from a variety of sources. So he just sent me a SQL dump of that and I was off running.

Measuring Similarity

The first question was how to measure the similarity between a given pair of tunes. Because the tune data online is typically stored in an ASCII format called ABC Notation, calculating string distances seemed the best solution. I tried both the Hamming and the Levenshtein distances and settled on Levenshtein because it supports comparing strings of different lengths.

Visualizing Similarities

I tried several ways of visualizing the melodies to highlight how few differences there were between pairs of tunes. Above is a more visually pleasing variant of my first attempts. I started by representing each note as a rectangle. Each row represents a tune. When a note is identical between that tune and the starting tune, the note is drawn. If the notes are different, that space is left blank. This gives you a picture of where the tunes are similar, but now how they differ.

My second attempt (the first image) overlays all tunes similar enough to the original. Again, each note is represented as a rectangle and only common notes are drawn. Each rectangle is mostly transparent in this visualization. Because all the tunes are drawn on top of each other, the darker the rectangle, the more often that note appears in that position in this set of tunes. Although I think this results in a rather attractive, barcode-like image, it doesn’t say much about how the tunes are similar.

Finally (at Golan’s suggestion) I represented the tunes as line graphs; plotting the notes in order horizontally and according to pitch vertically. Now you can see where the tunes overlap and how they differ when they don’t overlap. The important feature this rendition highlights is when a phrase is repeated between two tunes but out of phase with the original. Previously, common phrases were not visible unless they occurred in the same place in both tunes.

At this point I had to add some interactivity because simply plotting all the selected tunes on top of each other was too noisy to glean any useful information. So I faded out most of the lines and allowed users to select a single tune for comparison using the up and down arrows.

In the above side-by-side comparison, you can see that there are two identical tunes (the top pink bar on the left, and the top line chart on the right) with different names. A common occurrence in this style of music. The next most similar tune appears to have many similar phrases, but not in the same place as the original. Thus, on the left, it doesn’t appear to have much in common, but on the right you can see similar contours slightly out of phase with each other.

Critique

I think this is an interesting start, and the line graphs give me confidence that there are some interesting similarities to explore between tunes. I think a more interactive visualization might be more compelling; something in which users could explore the data more freely. The limited interactivity (see the video below) was mostly just a convenience for me to quickly generate artifacts for the presentation.

[vimeo=https://vimeo.com/36328177]

It’s also worth noting that my calculations of similarity are not correct. I simply treated the ABC data as strings, but that doesn’t result in a correct comparison because in many cases multiple characters are required to represent a single note. For example a low b is represented as “B,”, so the distance between the phrase “GB,D” and “GbD” should be one, because there is only one note—and thus one substitution—difference. However, simply comparing the strings “GB,D” and “GbD” results in a Levenshtein distance of two.

The fact that there are still so many similar looking tunes despite this oversight on my part is what encourages me that there is more to explore here.

Technical Details

I started out storing the data in a SQLite database, but it quickly grew too large and I moved it to a MySQL database. I used Python to perform the distance and longest common substring calculations on the tune data; that code can be found here. The visualization was performed in Processing, and that sketch can be found here. A zip file of the entire project—including a dump of the MySQL database—is available here (the SQL dump was too large to upload to WordPress).