This research proposes an automatic music transcription (AMT) method for converting vocal audio from “Shima-Uta,” a traditional folk music of the Amami Islands, into piano-roll-level notation that is accessible to contemporary audiences. Shima-Uta is undergoing cultural transformation as it transitions from its original community-based context to stage performances, even as younger generations remain actively involved. The transmission of this musical tradition is complicated by its reliance on oral transmission rather than established notation systems, as well as by the varying interpretations of individual songs among different performers and regional contexts. The distinctive characteristics of Shima-Uta, including extensive melismatic passages and unique vocal techniques, combined with limited available data, make it challenging to apply deep learning approaches that have shown superior performance in AMT. To address these constraints, we propose a hidden Markov model that captures frame-wise periodicity and top-down pitch variation biases derived from Shima-Uta scales, and employs the Viterbi algorithm to search for the optimal pitch sequence without requiring training data. Shima-Uta experts evaluated the transcription results inferred by our method and confirmed their effectiveness. Our approach facilitates documentation and cultural conservation of Shima-Uta through contemporary notation systems. Work supported by JSPS Grants 24H00715 and 24KJ1125.
Takahashi et al. (Wed,) studied this question.