MakamKit: A reliable tool for computational ethnomusicology research in Turkish music

Abstract

This article introduces MakamKit, an analysis engine developed for Turkish makam music, reports its measured performance, and discusses a transmission problem that surfaced during those measurements. The engine’s most fundamental property is that it carries the pitch world of makam music without distortion: reconstruction of comma information from digital scores was audited over 511 pieces and 129,151 notes and found virtually complete, with an error rate of two per thousand — the music is not rounded to twelve-tone equal temperament as it passes through the system. On this foundation, the engine identifies the makam of more than three quarters of pieces at first candidate on a 180-piece set it had never seen, rising to nearly nine in ten when restricted to the makam names it knows. Usul and form recognition are here measured numerically for the first time: across 2,586 pieces in seventeen usuls the correct usul appears among the top three candidates in about 93% of cases, and across 1,820 pieces in seven forms the correct form is identified at first candidate in about 87%. Makam recognition from audio is reliable in solo taksim, where the seyir is most exposed, and predictably more fragile in ensemble performances woven with modulations. In printed-score recognition, the engine computes which pitch a written sign corresponds to within a one-comma margin in nine notes out of ten, while it cannot reliably name the makam of that page — and rather than concealing this, it refuses to emit a clean score until expert verification is complete. This is the engine’s design principle: it stays silent when it does not know, ranks candidates when unsure, and declines entirely when the source is inadequate. That discipline yielded an unexpected finding: two digital copies of the same piece disagree about its final (karar) pitch in more than one case out of four. This is a problem of the corpus rather than of the engine, and it had not previously been measured at this scale; it shows that the digital copy is itself a link in the chain running from meşk to notation to print. The contribution is twofold: to introduce a research tool whose measured performance and source-trust architecture are documented openly, and to bring the transmission finding it enabled onto the field’s agenda. The software’s internal design is outside the article’s scope owing to its licence status.