On 13–14th of February 2025, I organised a #Hackathon on #RomaniChords that explored various methodological avenues for my #MSCA project’s data acquisition.
Here’s a short thread summarising what we dealt with and achieved during these days.
CCying everyone interested in #Ethnomusicology, #Musicology, #ComputationalMusicology, #MusicScience, #RomaniStudies, #RomaniMusic and alike.
Let’s jump right into it!
[Note: See alt text for more info about the pictures attached to this thread]
🧵1/20
The Hackathon brought together 10 scientists of different backgrounds – from linguists to music scholars and computer scientists.
To warm up intellectually,* we began with a general discussion of the phenomenon of Romani chords.
–––
*
Do you want to warm up with us to digest this thread easier? Watch the bit of the project’s video abstract from 7:17–7:58 and then try to answer the question: “Why does the dimenzovaný chord matter?”
=>
https://www.youtube.com/watch?v=aEXkI2br3cc&t=437s
🧵2/20
The discussion helped to identify the biggest challenges for the Hackathon:
On the one hand, the project aims to collect the data on an ambitious scale, which requires a pragmatic data-driven approach (a “musical laboratory”). But on the other, the phenomenon of Romani chords is embedded in peculiar layers of sociocultural reality, the uncovering of which requires a sensitive ethnographic approach.
This dilemma became the central theme for our methodological thoughts.
🧵3/20
Then, we discussed and tried out various contemporary music information retrieval methods with a special focus on accompanying #harmonies (aka #chords).
The first method to come up was “Automatic chord recognition” ( #ACR), a method that scientists have worked on for over 25 years.
A nice summary of this ongoing research (with the most significant challenges identified) can be found in this 5-year-old article => https://archives.ismir.net/ismir2019/paper/000004.pdf
🧵4/20
Nevertheless, we found the use of ACR for the #RomChords project to be limited.
ACR is a classification task that analyses audio signals to compare and categorise harmonic content into *predefined* chord labels.
While this method estimates very well the common progressions in functional harmony, it doesn’t quite know how to handle less common chords and voicings (such as the “dimenzovaný” chord” mentioned above) as they are not usually found in the *predefined* labels.
🧵5/20
Another method we considered was the #PolyphonicDetection.
This method focuses on identifying and extracting individual notes from musical textures by analysing frequency components, onset times, and timbral features using techniques such as spectral decomposition, pitch tracking, and machine learning-based transcription methods.
🧵6/20
