Paper Short Abstract:

This research applies Music Information Retrieval (MIR) and machine learning methods to automatically cluster instrumental folk music recordings based on their melodic and harmonic content. In addition to identifying different performances of the same musical piece, this approach also facilitates the detection of errors within digital archives, such as misattributed recordings or duplicate entries.

Paper Abstract:

Ethnomusicological analysis of folk music traditionally relies on the comparison of transcriptions derived from audio recordings, rather than analyzing the audio itself. This approach necessitates the manual transcription of recordings, a process that is both time-intensive and impractical for large datasets. The Estonian Folklore Archive houses an extensive repository of instrumental folk recordings, many of which have their potential interconnections yet to be uncovered. Additionally, there is a growing need to identify inconsistencies within digital archives, such as misattributed recordings or duplicates.

The task of identifying different performances of the same musical piece—commonly referred to as cover song identification in the field of Music Information Retrieval (MIR)—has seldom extended beyond manual approaches in ethnomusicology. MIR combines musicology, computer science, and signal processing to extract musically relevant information directly from audio signals. Despite its widespread use in popular music (e.g., for classification or recommendation systems), its applicability to archival recordings remains largely unexplored.

In my Master’s thesis, I leveraged MIR and machine learning to automatically cluster instrumental folk music recordings based on their melodic and harmonic content, addressing the challenge of identifying “covers”—i.e., different performances of the same piece—and detecting anomalies within the archives. This presentation discusses the findings of this research and hopes to demonstrate the potential of MIR methods to uncover patterns and relationships not immediately discernible through traditional manual methods. While not intended to replace manual analysis, computational approaches offer a valuable complementary tool in ethnomusicological studies to enhance the efficiency of analyzing, comparing, and classifying audio recordings.