About & Sources
The academic resources, corpus data, and open-source computational systems that power Mithlond's Elvish speech pipeline.
The Computational Philology & Analytics Process
Behind Mithlond's speech synthesis lies a rigorous linguistic analysis engine that digitizes historical Elvish philology. By parsing J.R.R. Tolkien's massive corpus and Paul Strack's comprehensive lexicon, we reconstruct etymological trees, map sound mutations over fictional millennia, and analyze the mathematical phonotactics of Elvish word construction.
Phase 1
Lexical Ingestion
Eldamo Schema Parse
Extracting complex vocabulary hierarchies, parts of speech, and developmental periods from Eldamo's extensive historical lexicon data dumps.
Phase 2
Sound Law Mapping
Diachronic Rules
Modeling the chronological sound shift rules that transformed Primitive Common Eldarin roots into distinct Quenya and Sindarin dialects.
Phase 3
Etymology Graph
Network Reconstruction
Mapping root-and-derivative family trees into dynamic, interconnected nodes to visualize semantic inheritance and vocabulary evolution.
Phase 4
Phonotactic Analysis
Bigram & Morphology
Running character transition matrices, bigram cluster modeling, and morphological sunburst aggregations to analyze structural trends.
The Mithlond Speech Synthesis Pipeline
Mithlond reconstructs spoken Tolkien languages by bridging historical philological databases with modern neural text-to-speech models. Text undergoes a series of complex transformations: from root verification to syllabification, phoneme transcription, phonetic alignment, and finally expressive neural wave synthesis.
Step 1
Lexicon Lookup
Eldamo Database
Tolkien's vocabulary and inflections are cross-referenced with the Eldamo database to verify morphological properties and etymological roots.
Step 2
Syllabification
Arda Library
Digital Tolkien's Arda library applies precise rules of syllable boundaries, vowel lengths, and consonant mutations to isolate syllables and accent stress.
Step 3
G2P Transcription
Custom G2P Engine
Mithlond's grapheme-to-phoneme engine translates syllabified Tolkien representations into standard IPA strings and mappers compatibility.
Step 4
Neural Audio Synthesis
kokoro-cli & Model
The high-performance Rust kokoro-cli runs the open-weight Kokoro TTS model, synthesizing expressive WAV streams at sub-50ms speeds.
Eldamo Database
Eldamo (Elvish Data Model) is the premier historical lexicographical database of Tolkien's languages, spanning Quenya, Sindarin, Primitive Elvish, Telerin, Gnomish, Noldorin, and more. It documents the diachronic development of Tolkien's linguistics, cataloging roots, inflections, historical sound shifts, and vocabulary evolution. Mithlond utilizes Eldamo data dumps as its primary data model.
Visit eldamo.org
Arda Python Library
The Arda library is an open-source tool developed by the Digital Tolkien Project. It implements J.R.R. Tolkien's precise linguistic logic for Quenya and Sindarin. Arda specializes in proper syllable division, vowel length weightings, vowel/consonant combination mutations, and locating standard word stress to establish authentic phonetic shapes.
Digital Tolkien Project
G.H. Chinoy's Pronunciation Tools
A suite of custom Elvish pronunciation tools developed specifically to bridge Tolkien's complex linguistic rules with neural TTS pipelines. It integrates the Arda library, applies syllable weight analyses, handles custom Grapheme-to-Phoneme (G2P) transcription maps for Quenya and Sindarin, and serves the live API backend powering Mithlond's audio capabilities.
Custom Implementation
kokoro-rs (Rust TTS Engine)
A unique, high-performance native Rust implementation and CLI tool designed specifically to run the Kokoro text-to-speech neural model. By executing ONNX Runtime inference directly inside compiled Rust, it completely eliminates Python interpreter overhead, resulting in blistering sub-50ms audio stream generation suitable for interactive web environments.
View My kokoro-rs Repository
Kokoro TTS Model
Kokoro is a state-of-the-art, open-weight text-to-speech neural model featuring 82 million parameters. Despite its highly compact size, it delivers exceptionally expressive, warm, and natural-sounding audio quality. Because it relies heavily on direct phonetic (IPA) input vectors for its synthesis generation, it is uniquely suited for custom linguistic programming, allowing us to feed custom Elvish IPA directly into its neural voice nodes.
hexgrad/kokoro on GitHub