Voces ex Machina: A Corpus of Prosodically-Correct Latin Poetic Audio from Large-Language-Model TTS

# Voces ex Machina: A Corpus of Prosodically-Correct Latin Poetic Audio This repository contains the dataset for the paper "Voces ex Machina: A Corpus of Prosodically-Correct Latin Poetic Audio from Large-Language-Model TTS" by Michele Ciletti. This corpus provides openly licensed, prosodically-correct audio recordings of classical Latin poetry generated using a large language model (LLM) text-to-speech (TTS) engine. The goal is to make the quantitative rhythm of Latin verse, which is based on the alternation of long and short syllables, audibly accessible to students, researchers, and the general public. The dataset includes 216 lines of poetry from Vergil's *Aeneid* and Ovid's *Heroides*, totaling approximately 24 minutes of audio. Each line is provided as an individual audio file and as part of a continuous recitation. The audio is aligned with several layers of textual and metrical annotation, making it a valuable resource for pedagogy, accessibility, computational linguistics, and speech synthesis research. ## Corpus Details | Sub-corpus | Metre | Lines | Hexameters | Pentameters | Total Duration | | :--- | :--- | :---: | :---: | :---: | :---: | | **Aeneid 1.1–100** | Dactylic Hexameter | 100 | 100 | 0 | 00:11:26 | | **Heroides 1.1–116**| Elegiac Couplet | 116 | 58 | 58 | 00:12:26 | | **Total** | | **216** | **158** | **58** | **00:23:52** | ## Dataset Contents The repository is organized as follows: ``` . ├── aeneid_audio/ │ ├── full.mp3 │ ├── line_1.mp3 │ ├── line_2.mp3 │ └── ... (up to line_100.mp3) │ ├── heroides_audio/ │ ├── full.mp3 │ ├── line_1.mp3 │ ├── line_2.mp3 │ └── ... (up to line_116.mp3) │ ├── aeneid_metrical_scansion.xml ├── heroides_metrical_scansion.xml │ ├── aeneid_original.txt ├── heroides_original.txt │ ├── aeneid_pre_processed.txt └── heroides_pre_processed.txt ``` ### File Descriptions * **`*_audio/`**: These directories contain the generated speech. * `full.mp3`: A single audio file with the continuous recitation of the entire passage. Lines are separated by a uniform 800ms pause. * `line_###.mp3`: Individual audio files for each verse, allowing for isolated listening or analysis. * **`*_metrical_scansion.xml`**: The metrical information for each line, exported from the [Pedecerto](https://www.pedecerto.eu/) project. This XML file includes syllable divisions, vowel quantities, ictus (metrical stress) placement, and caesura locations. * **`*_original.txt`**: The original Latin text of the passages in standard classical orthography. * **`*_pre_processed.txt`**: The version of the text that was fed to the TTS model. This text has been orthographically "nudged" to guide the model towards a more accurate pronunciation. This is the direct input that generated the corresponding audio. ## Methodology The creation of this corpus followed a multi-step process to ensure metrical fidelity. 1. **Source Texts & Scansion**: The first 100 lines of Vergil's *Aeneid* and 116 lines of Ovid's *Heroides* 1 were chosen. Metrical scansion for each line was sourced from the Pedecerto project's XML exports. 2. **Text Pre-processing**: To guide the TTS model, each line of verse was automatically processed. This "TTS-ready" text is what you find in the `*_pre_processed.txt` files. The key transformations were: * **Ictus Marking**: Syllables receiving metrical ictus were capitalized. The vowel within that syllable received a grave accent (e.g., `arma` -> `ÀRma`). * **Elision**: Compulsory elisions were resolved graphically (e.g., `quoque et` became `quoquet`). * **Orthographic Adaptation**: Graphemes that often confuse English-trained TTS models were replaced with more phonetically transparent equivalents. | Original Grapheme | TTS-Ready Grapheme | Example | | :--- | :--- | :--- | | `c` (before e, i) | `k` | `cano` -> `kano` | | `qu` | `kw` | `qui` -> `kwi` | | `g` (before e, i) | `gh` | `regina` -> `reghina` | | `ae` | `ai` | `Troiae` -> `Troiai` | | `oe` | `oi` | `poenae` -> `poinai` | 3. **Speech Generation**: The pre-processed text for each line was fed to OpenAI's `gpt-4o-mini-tts` model using a specific prompt designed to enforce a rhythmic, articulated delivery. The system prompt was: ``` 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] ``` 4. **Human Validation**: Ten audio versions of each line were generated. Experts in Latin phonology audited every recording for segmental accuracy, correct stress placement, elision, and pacing. Only the highest-scoring, error-free rendition for each line was accepted into the final corpus. 5. **Audio Mastering**: The selected audio files for each verse were loudness-normalized and concatenated with a uniform 800ms silence between them to create the `full.mp3` recitations. ## Use Cases This dataset is designed for a variety of purposes: * **Pedagogy**: A classroom resource for students to hear the natural rhythm of Latin epic and elegiac poetry. * **Accessibility**: Provides an auditory version of classical texts for visually impaired learners or those who learn best by listening. * **Research**: A test-bed for developing and evaluating prosody-aware speech synthesis models. * **Fine-Tuning**: The aligned audio and text can serve as a small but high-quality dataset for fine-tuning TTS models for Latin or other quantitative-metre languages. * **Reproducibility**: Researchers can use the `*_pre_processed.txt` files to attempt to replicate or extend the results with different TTS models. ## License All components of this dataset (audio, text, and annotations) are released under the [Creative Commons Attribution 4.0 International (CC BY 4.0) license](https://creativecommons.org/licenses/by/4.0/). You are free to share and adapt the material for any purpose, even commercially, as long as you give appropriate credit. ## Acknowledgments * The [Pedecerto](https://www.pedecerto.eu/) team for their monumental work in creating and sharing machine-readable Latin scansions. * The [Classical Language Toolkit (CLTK)](https://cltk.org/) community for open-source tools that aided in syllabification. * Colleagues at the University of Foggia for their time and expertise in the auditory validation of the recordings.