CLiC-it 2025 · 2025 · Workshop
Michele Ciletti · → Paper · Demo: ✓ · Code: ✓
Introduces an openly licensed, line-aligned corpus of 216 expert-validated Latin poetic audio recordings (approximately 24 minutes) generated via prompt-engineered GPT-4o-mini-tts, covering Vergil’s Aeneid (100 hexameters) and Ovid’s Heroides (116 elegiac couplets), with metrical annotations drawn from the PedeCerto project.
Problem
Classical Latin metre organises verse around the alternation of long and short syllables (vowel quantity), not stress as in most modern poetry. High-quality recordings that faithfully reproduce this quantitative rhythm are rare and costly to produce with trained classicists. General-purpose TTS models, trained predominantly on modern languages, fail to preserve long-short alternation and instead project the stress patterns of English or Romance languages onto Latin text, producing rhythmically inaccurate speech unsuitable for pedagogy or computational research. The resulting absence of quality reference material leaves students without audible models and speech technologists without a test bed for prosody conditioning in low-resource historical languages.
Method
The workflow combines orthographic pre-processing, prompt-engineered LLM synthesis, and expert human validation.
Pre-processing begins with machine-readable metrical scansion exported from the PedeCerto project XML. Each verse is syllabified using the Classical Language Toolkit (CLTK) rule-based module. Ictic (metrically stressed) syllables are marked with grave accents and converted to capitals. Obligatory elisions are resolved graphically (for example, quoque et becomes quoquet). Commas are inserted at caesura positions unless the manuscript already supplies punctuation at that location. A second pass applies grapheme substitutions to reduce confusion in English-trained acoustic models: 〈c〉 becomes 〈k〉 before front vowels, 〈qu〉 becomes 〈kw〉, 〈ae〉 and 〈oe〉 become 〈ai〉 and 〈oi〉, and palatal 〈g〉 expands to 〈gh〉. Each line is synthesised in isolation to keep prosodic control local.
The pre-processed line is passed to GPT-4o-mini-tts via the OpenAI API, with a system prompt instructing the model to recite slowly, articulate every syllable, and obey the marked stresses. Ten readings are requested per line. The final system prompt template is: “This is a Latin poetical verse. Pronounce it rhythmically, slowly and with emphasis, articulating each syllable and correctly stressing them. Pronounce it like this: [pre-processed verse].”
Specialists in Latin phonology audit all ten candidates per line, classifying errors as segmental substitutions or ictus misplacements. Acceptance requires no stress or elision error and no more than minor segmental deviation. The single best rendition per line is loudness-normalised and concatenated with 800 ms inter-verse silences to produce two continuous recitations (one per poem). All assets are deposited on Zenodo under CC-BY 4.0.
Key Results
The workflow produced at least one acceptable recording for each of the 216 metrically autonomous lines, yielding approximately 24 minutes of expert-validated Latin poetic audio (11:26 for the Aeneid excerpt, 12:26 for the Heroides excerpt). No standard TTS evaluation metrics (MOS, WER, etc.) are reported; quality assessment is binary expert accept/reject.
Orthographic substitutions reduced segmental substitution errors by approximately one third without any model retraining (§5.2). Caesura commas proved particularly effective for enforcing rhythmic pauses at obligatory pentameter mid-verse positions. Cross-lingual phonological interference was consistently observed: passages with Romance vocabulary received Italian-style intonation, English loanwords triggered anglophone timbre, and -rant word endings sometimes produced audibly Iberian cadences. This confirms that Latin-specific data is essentially absent from the model’s training corpus.
Novelty Assessment
The contribution is a small but carefully assembled corpus rather than a technical advance in speech synthesis. The methodology assembles existing components: PedeCerto scansion, CLTK rule-based syllabification, a proprietary API (GPT-4o-mini-tts), and hand-crafted orthographic substitution rules. No architecture is proposed, no model is trained, and no new TTS conditioning mechanism is introduced. The orthographic adaptation heuristics are ad hoc rather than learned, and the paper acknowledges this explicitly. The key transferable observation, that prompt engineering and input preprocessing can partially compensate for the absence of low-resource language training data, is stated clearly and supported by systematic human validation, giving the workflow value as a replicable template for other historical or minoritised verse traditions.
Field Significance
Low — the paper provides a narrow-domain open corpus (216 lines of classical Latin verse, approximately 24 minutes) assembled via prompt-engineered LLM TTS. The workflow can serve as a template for similar efforts in Ancient Greek, Old Occitan, or other verse traditions with existing digital scansion, but the corpus is too small for TTS fine-tuning, relies on a proprietary API, and addresses a specialised domain that falls outside mainstream TTS benchmarking.
Claims
-
supports: Orthographic pre-processing of input text (accent marking, grapheme substitution, elision resolution) can reduce segmental substitution errors in LLM-based TTS applied to low-resource or historical languages without any model retraining.
Evidence: Replacing rare digraphs and adapting graphemes (〈c〉 to 〈k〉, 〈ae〉 to 〈ai〉, 〈qu〉 to 〈kw〉) reduced segmental substitutions by approximately one third; no retraining of GPT-4o-mini-tts was performed. (§3.2, §5.2)
-
supports: Natural language prompt instructions can steer general-purpose LLM-based TTS toward prosodically-correct speech in low-resource languages without fine-tuning.
Evidence: A concise system prompt specifying slow pace, explicit stress, and syllable articulation, combined with orthographically pre-processed input, produced at least one expert-validated recording for all 216 Latin verse lines. (§3.3, §3.4)
-
complicates: LLM-based TTS systems trained on modern multilingual data produce systematic cross-lingual phonological interference when applied to historically or orthographically distinct languages.
Evidence: GPT-4o-mini-tts imposed Italian, English, or Spanish prosodic templates on Latin text depending on vocabulary similarity; the phenomenon confirms that Latin-specific data is essentially absent from the model’s training corpus. (§5.1)
-
complicates: Prompt-based prosody control is more effective for stress placement than for enforcing absolute phonemic length distinctions.
Evidence: The workflow approximated Latin vowel quantity through slower pacing on ictic syllables but could not enforce a fixed heavy-to-light duration ratio, the standard for quantitative metre in phonetic work. (§5.6, §6)
Limitations and Open Questions
Warning
The corpus depends on GPT-4o-mini-tts, a proprietary model with inaccessible weights. Should the API access policy change, exact reproduction of the dataset becomes impossible.
The corpus is small (216 lines, approximately 24 minutes), falling well short of the material typically required for full end-to-end TTS fine-tuning. Stylistic range is narrow: both texts belong to the Augustan literary period and formal register; colloquial Latin, post-Classical spelling, and late-antique forms are untested. A single synthetic voice is used throughout, precluding multi-speaker or gender-diverse analysis. Only the classical reconstructed pronunciation is represented; ecclesiastical pronunciation is excluded. Expert validation, while rigorous, introduces selection bias from a single institution’s phonological tradition. Quantity is approximated through pacing rather than explicit duration control; no fixed heavy-to-light ratio is enforced, leaving a gap relative to the phonetic gold standard.
Wiki Connections
- Prosody Control — introduces an explicit mechanism (ictic accent marking, grapheme substitution, caesura punctuation, prosodic prompt) to control quantitative metre and stress in Latin TTS without model retraining
- Instruction-Conditioned TTS — demonstrates that a natural language system prompt specifying pace, articulation, and stress can successfully condition GPT-4o-mini-tts toward metrically faithful Latin speech
- Subjective Evaluation — uses domain-expert listeners (Latin phonologists) to validate recordings with binary accept/reject judgements, classifying errors as segmental substitutions or ictus misplacements