Gomanize

A Go library and CLI that romanizes Devanagari (Hindi) into readable Latin script, built for song lyrics and other colloquial text. Rule-based engine with optional embedded learned components; no runtime dependencies.

नमस्ते भारत  →  namaste bharat

Try it in your browser → budhash.com/gomanize — the full engine, compiled to WebAssembly, runs entirely client-side (no server; no text leaves your machine).

It does romanization — spelling Hindi the way it sounds (नमस्ते → namaste) — rather than the strict, reversible transliteration of schemes like IAST or ISO 15919. ("Transliteration" is the common colloquial term for the same task, and the API keeps the Translit name.)

Romanization has no single correct answer — जनता is validly janata, janta, or janataa — so this project measures itself against all human-attested variants. On its primary use case, romanizing verse, its character error rate is at the level where human romanizers disagree with each other. Details, methodology, and known limitations: docs/RESEARCH.md.

Install

# Library
go get github.com/budhash/gomanize

# CLI from source
git clone https://github.com/budhash/gomanize
cd gomanize
make build

Usage

CLI

./gomanize "नमस्ते भारत"        # namaste bharat
echo "हिंदी गाना" | ./gomanize   # hindi gana

Flag	Effect	Example
(default)	Colloquial rules	जनता → janta
--keep-medial-schwa	Retain medial schwa	जनता → janata
--long-vowels	aa for every ā	गाना → gaanaa
--simple-nasals	Simplified nasal endings	करें → karen
--schwa-model	Learned schwa classifier	जनता → janta
--lexicon	Attested spellings for 8,367 known words	अंकल → uncle
--rerank	Character-LM picks best of rules/schwa-model outputs	see Accuracy below
--list-rules, --debug	Inspect and trace the rule engine

Library

import gomanize "github.com/budhash/gomanize"

g, err := gomanize.New("hindi")
if err != nil {
    panic(err)
}
fmt.Println(g.Translit("नमस्ते दुनिया")) // "namaste duniya"

Options mirror the CLI flags via gomanize.NewWithOptions.

Accuracy

Scored against all human-attested romanization variants (see docs/RESEARCH.md for methodology, datasets, and the full result set including negative results):

Benchmark	Result
Curated Dakshina, default rules	92.8% (94.7% with --rerank)
Naturally-typed Hindi (COMI-LINGUA), token-weighted	78.9% (85.7% with --lexicon)
Held-out unseen words	69.0% (70.4% with --rerank)
Song lyrics, line-level character error	0.049, or 0.039 with --lexicon (human agreement floor is about 0.054)

Known limitations:

• Vowel-length spelling (ee/oo vs i/u) is a lexical convention, not rule-governed; roughly half of remaining errors are this class. Positional rules for it measured net-negative and were rejected (docs/RESEARCH.md, §5).
• Named entities and loanwords are weak without --lexicon (names have conventional spellings no phonetic rule can produce: दीक्षित → dixit).
• Different annotation efforts use different conventions; scores against corpora with other conventions (e.g. Aksharantar's frequent-word slice) are substantially lower than against Dakshina-style references.

Documentation

Document	Contents
docs/RESEARCH.md	The problem, literature, datasets and licenses, evaluation methodology, all results including negatives
docs/DESIGN.md	Engine architecture, rule system, character mappings, learned components, future directions
docs/ROADMAP.md	Post-1.0 directions (convention schemes, more languages) with tradeoffs
web/README.md	The browser demo (budhash.com/gomanize): make wasm / make wasm-serve and the Pages deploy
CLAUDE.md	Development workflow, commands, repository conventions
docs/PROCESS.md	Task tracking, PR discipline, accuracy reporting rules
docs/reviews/	Dated decision records for every result, including failures

Development

make init     # First-time setup (deps + pre-commit hooks)
make ci       # Full pipeline: format check, lint, build, tests, benchmarks
make help     # All commands

Contributions follow docs/PROCESS.md: feature branches, make ci before PRs, accuracy changes must show before/after on the benchmark suite.

License

MIT. Copyright (c) 2023-2026 Budhaditya (budhash@gmail.com).

Benchmark data derives from Dakshina (CC BY-SA 4.0), Aksharantar (CC-BY 4.0), COMI-LINGUA (CC-BY 4.0), and Shabd (CC0); see docs/RESEARCH.md for full attribution.