Prompting the Muse

Generating Prosodically Accurate Audio of Latin Poetry with Text-to-Speech Large Language Models: A Computational Workflow

Abstract

While the field of Digital Humanities has successfully established robust infrastructures for the textual analysis of Latin, the auditory dimension of the language is still largely undeveloped. Specifically, the complex quantitative rhythm and intonation of classical poetry cannot be accurately replicated by Text-to-Speech models. This paper presents a computational workflow designed to bridge this gap, leveraging verified metrical data to produce high-fidelity and prosodically accurate audio recordings. By using the structured XML scansions of the Pedecerto project, the proposed pipeline employs a rule-based pre-processing routine to convert standard orthography into a phonetic script optimized for acoustic modelling. These adapted texts are then fed into a multimodal Large Language Model, which is steered via in-context prompt engineering to observe syllable quantity, ictus placement, pauses, and eventual elision. The technical architecture of this system is detailed, analyzing the specific orthographic interventions and prompts required to overcome the stress-timed bias of contemporary AI models. Finally, the implications of this tool for the wider Digital Humanities ecosystem are discussed, with particular attention to its potential to democratize access to Latin learning, support accessibility, and add new audio layers to existing digital projects and infrastructures.