Prompting the Muse: Generating Prosodically-Correct Latin Speech with Large Language Models

This study investigates whether prompt engineering, aided by symbolic prosodic markup, can induce a large language model-based text-to-speech (TTS) system to recite Latin verse with metrically correct stress. Two hundred dactylic lines—one hundred hexameters from Vergil’s Aeneid and the opening elegiac couplet sequence of Ovid’s Heroides—were taken from the Pedecerto corpus, whose XML already encodes syllabification, ictus positions, and caesurae (Colombi et al., 2011). A preprocessing pipeline syllabified each verse, merged obligatory elisions, inserted commas at caesurae, transliterated graphemes toward an English-Italian compromise, upper-cased every ictic syllable, and placed a grave accent on its vowel. The verses were supplied to gpt-4o-mini-tts (Hurst et al., 2024), which resulted to be the best-performing model for the task, through a compact system prompt requesting slow, articulated speech that obeys the visible stresses. Ten stochastic realisations per line were generated; three independent Latin scholars evaluated the output, labelling segmental, stress, elision, and pacing errors. At least one recording met all criteria for 91% of the 216 lines. Ablation confirmed that case-shift plus accent carried most of the improvement, whereas syllable hyphenation was superfluous. Persistent errors clustered around lexical cognates whose default English or Romance stress conflicts with the metrical ictus. The present workflow achieves a promising prosodic fidelity while leveraging only text-level cues and the model’s general acoustic decoder, following an approach similar to the one outlined by Lam et al. (2025). It also sidesteps the scarcity of duration-marked training data that hampers quantitative reconstructions of Latin phonology (De Sisto et al., 2024). The resulting corpus, which consists of aligned text, validated audio, and scripts, will be openly released on Zenodo, supporting pedagogy, accessibility, and computational metrics research. Future work aims to couple the current strategy with controllable duration predictors offered by FastSpeech-type models (Ren et al., 2020) and train lightweight alignment modules to reduce the need for manual best-of-N selection, in line with broader calls for FAIR audio resources in Digital Humanities (De Sisto et al., 2024).

Nov 5, 2025 11:30 AM — 12:30 PM