AI-based Analysis of Gayageum Sanjo

Minji Kim(First Author), Hongkoo Yeo(Co-Author), Dasaem Jeong(Corresponding Author)

Libraries: numpy, pandas, matplotlib, pytorch, scipy

This paper explores how artificial intelligence and audio signal processing can be used to analyze and visualize nonghyeon—the expressive vibrato technique central to Korean traditional instrumental performance—using the case of gayageum sanjo.

Co-authored as an academic research paper, the study applies a deep learning-based pitch extraction model (CREPE) alongside Fast Fourier Transform (FFT) analysis to quantitatively examine nonghyeon at the level of sobak (the smallest rhythmic subdivision within a jangdan cycle). By extracting precise pitch contours from recorded performances and analyzing their periodicity, depth, and rate, the research reveals consistent structural patterns in vibrato execution that persist across performers, string positions, tonal functions, and phrasing contexts.

A key finding is that nonghyeon is not merely an expressive ornament but is systematically aligned with rhythmic structure: for example, approximately three vibrato cycles per sobak in jinyangjo and two in jungmori were consistently observed. These results challenge the notion that nonghyeon exists solely as an intuitive or subjective technique transmitted through oral tradition and instead demonstrate that it possesses measurable, repeatable temporal characteristics.

Beyond analysis, the project emphasizes visualization as a pedagogical tool. By transforming traditionally implicit performance knowledge into interpretable pitch curves and frequency-domain representations, the study proposes a framework for AI-assisted feedback in Korean traditional music education. This approach aims to lower the learning barrier for students, support reflective practice, and bridge embodied musical knowledge with computational methods.