Can a Machine Learn Thai Musical Grammar? Symbolic Modeling of Thai Classical Music

What happens when computational methods encounter a musical tradition whose knowledge has largely been transmitted orally? This paper presents a symbolic framework for preserving, analyzing, and generating Thai classical music. We construct a machine-readable corpus of 212 songs and 1,243 sections from scanned notation while retaining the tradition’s hierarchical tier, section, and bar structure. Using pitch, interval, cadence, and phrase-pattern features, we train models to classify six culturally associated motif families. The results reveal musically meaningful relationships: Khmer repertoire emerges as particularly distinctive, while overlap between Burmese and Mon styles reflects historical processes of musical absorption. We then investigate whether generative models can learn Thai melodic grammar. N-gram and conditioned LSTM baselines reproduce local stylistic patterns but struggle with long-range phrase return. Explicit rhythmic constraints substantially reduce invalid note placement, demonstrating how cultural knowledge can be encoded directly into model behavior. Rather than presenting AI as an autonomous composer, this research uses generation as a form of inquiry: what can a machine learn from a small, culturally specific corpus, and where does its understanding fail? The resulting Sepha.ai prototype connects the research to practice through searchable notation, playback, similarity retrieval, and experimental generation. The project argues that computational modeling can support preservation and education while keeping musical interpretation and authority with human practitioners.