It has been a while now that the two authors of this particular academic blog have been interested in computational methods applied to authorship attribution. Until now, our research have been focused on texts, mostly from the Middle Ages and the XVIIth century. Given our shared scientific interest for digital musicology, one question arised: can our methods, designed for textual sources, apply to scores ? Finding out which compositor wrote a specific piece of music seems even more tricky than finding the author of text.
Looking for fingerprints
To understand who could be a text’s author, we know which clues to look for. Most of the time, the author of a text will reveal his identity through the unconscious repetition of certain grammatical structures, the use of certain function words, or, for versified texts, the choice of a specific vocabulary for his rhymes.
What would be the equivalent of all these phenomena in a musical work ? Is it possible to define a short list of specific phenomena that we should spot to identify a composer’s signature, no matter the style or the era ?
Here are a few insights gained from this year’s edition of the Digital Humanities at Oxford Summer School.
Are improvisations signatures ?
Using material from Stanford University’s Josquin Research Project (ca. 1420- ca 1520) (Stanford University) and from the CESR (Tours) Les livres de Chansons Nouvelles de Nicolas Du Chemin (1549–1568) project, Catherine Motuz gives us a first possible answer.
During the the fifteenth and sixteenth centuries, musicians were used to improvise codified and sometimes very complex polyphony. The structures of these improvisations follow very strict rules, recorded in various treatises by several authors, from Guillelmus Monachus around 1480 to Tomàs de Santa Maria in the middle of the 1560s. Yet, these rules leave a composer with a few choices, and each composer seems to favour specific formulas, that could give his identity away.
To work on the structure of these polyphony, one have to study the succession of vertical intervals – which Catherine Motuz does through the Electronic Locator of Vertical Interval Successions (ELVIS). Working on polyphonic n-grams (yes, musicologists use n-grams for authorship attribution too!), she detects the use of different polyphonic models – canons, parallel sixths, parallel tenths… With two-grams, the separation between Josquin and Du Chemin is not obvious ; but with three-grams, the contrast is striking!
Tell me how you end, I will tell you who you are
Another work by Frauke Jurgensen points in the same direction. Studying chansons written by composers from the “Burgundian school”, mostly by Gilles Binchois (ca 1400- 1460) and Guillaume du Fay (1397-1474), she tries to identify unique features of those composers. Binchois writed tunes more jagged than du Fay. He also differentiated more among voices than Du Fay, in terms of voice leading style. One place though looks like a reasonable candidate to look for specificities: cadences.
Even if identifying a cadence, especially when they are somewhat elaborate, can be tricky for the computer, Jurgensen manages to clearly isolate 8000 cadences. Their study shows that, even if there is no such thing as a cadence specific to Binchois or du Fay, big differences can be exhibited between the way these two authors write their cadences. They mostly use the same patterns, but not with the same frequency. Computing the proportion of certain cadences in an unattributed piece could then help find its plausible author.
What is next ?
Following these two studies, it seems that authorship attribution in musicology will not easily be able to reproduce the methodology used in text studies. For now, only an expert view on the compositions from a certain place and period might allow to identify the markers of a certain composer. To understand who is the author of an unattributed piece of music, we would have to spot a certain feature in this piece, different according to the genre, place and time, and compare it to the way this feature is normally executed by one or the other composer.
Will we find techniques that we could apply more generally ? The properly digitized corpora are still rare, and the studies in this field look quite rare (I would be grateful to any reader pointing me to papers addressing this problem). It is thus very difficult to answer this question – but I am eager to find out !
ANTILA Christopher, CUMMING Julie, “The VIS Framework: Analyzing Counterpoint in Large Datasets”, in ISMIR (pp. 71-76), 2014
JURGENSEN Frauke, “Symbolic Music Analysis of Renaissance Counterpoint: current challenges”, Digital Humanities at the Oxford Summer School, Oxford University, July 5th 2017
MOTUZ Catherine, “Authorial fingerprinting for Renaissance Polyphony: Finding Contrapuntal Formulas in Two Corpora of French Chansons”, Digital Humanities at the Oxford Summer School, Oxford University, July 4th, 2017