LIPI — compact Hindi Devanagari → Roman transliteration

LIPI romanizes Hindi the way people actually type it — natural "Hinglish" spellings for messaging, search, dictation, and social media — with a 534,314-parameter non-autoregressive akshara-CTC model built for CPU and edge inference.

नमस्ते दुनिया   →   namaste duniya
कैसे हो दोस्त   →   kaise ho dost

A Hindi-only model ~20× smaller than the standard multilingual baseline (IndicXlit, 11 M), LIPI matches it on standard transliteration and is far stronger on naturally typed Hindi.

Model details

Property	Value
Parameters	534,314
Architecture	codepoint-composed akshara encoder → learned slot expansion → 2-layer BiGRU → CTC
Source / target vocab	87 / 42
Max codepoints per cluster	16 · slots per akshara 6 · model dim 96 · GRU hidden 128×2
Decoding	greedy CTC (default); optional prefix beam
Size	fp32 2.2 MB · dynamic int8 ~0.57 MB
Framework	PyTorch (custom runtime, not transformers)
SHA-256	ea5013c9… · seed 20260814 · PyTorch 2.6.0 / Python 3.11

Usage

LIPI ships a small runtime (the lipi package), not a transformers pipeline.

pip install torch huggingface_hub
pip install git+https://github.com/ABHISHEKgauti25/lipi.git

from huggingface_hub import snapshot_download
from lipi import Transliterator

model_dir = snapshot_download("abhishekgautamm/lipi")   # weights + configs
t = Transliterator(model_dir=model_dir)                  # backend="lipi"

t.transliterate_text("नमस्ते दुनिया")     # "namaste duniya"
t.transliterate_word("एक्सप्रेस")         # "express"

Benchmarks

Standalone LIPI vs. IndicXlit (11 M) and ITRANS (rule-based) on held-out test splits, one consistent harness. Higher is better for accuracy; lower for CER.

Benchmark (test)	Metric	LIPI	IndicXlit	ITRANS
FIRE 2013 (typed search)	exact	72.48	34.83	43.84
Aksharantar	exact	40.73	39.67	11.82
Dakshina words	in-attested	77.76	73.04	26.68
Dakshina sentences	word-match	65.63	32.97	30.07
IndoNLP 2025 (Set 1)	word-match	66.24	32.96	31.68
FIRE 2013	CER ↓	8.31	21.17	20.32

Dev + test with all metrics: see the GitHub benchmarks.

Training data & evaluation integrity

LIPI is trained on a usage-ranked lexicon built train-only from open resources (Aksharantar train, Dakshina train lexicon, Xlit-Crowd), ranked by attestation plus a Roman-usage prior from L3Cube-HingCorpus (train split). No development or test material from any benchmark is used in training, and dataset-intrinsic train↔test vocabulary overlap is ≈ 0 — so the numbers above are genuinely held out.

Optional: lexicon-first deployment

This repo also ships lipi_map.json (~45 MB), a curated exact-lookup lexicon. For deployments that want frequent forms canonicalized at full coverage, use the map_lipi backend (lexicon first, LIPI on a miss):

t = Transliterator(backend="map_lipi", model_dir=model_dir,
                   map_path=f"{model_dir}/lipi_map.json")
t.transliterate_text("मैंने कहा कि भारत महान है")   # "maine kaha ki bharat mahan hai"

The lexicon is training-data-derived and is not part of LIPI's reported benchmarks.

Limitations

• Multiple Roman spellings can be valid (nahi / nahin); strict exact-match undercounts.
• Schwa deletion and context-dependent forms are hard; transliteration is per-token.
• Proper names, rare words, and borrowings may romanize unexpectedly.
• Output style reflects the training distribution and annotator preferences. Not a formal ISO/IAST/ITRANS romanizer, a translator, or a Roman→Devanagari model.

License

CC BY-NC-SA 4.0 (NonCommercial, ShareAlike). LIPI is trained on data derived from Xlit-Crowd (CC BY-NC-SA 4.0) and Dakshina (CC BY-SA 4.0), so the weights and the lexicon inherit these terms. For commercial use, retrain on commercially-licensed data. Please attribute Dakshina, Xlit-Crowd, Aksharantar, and L3Cube-HingCorpus. The accompanying source code is Apache-2.0.

Citation

@software{lipi_2026,
  title  = {LIPI: A Compact Devanagari-to-Roman Transliteration Model for Natural Hindi},
  author = {Gautam, Abhishek},
  year   = {2026},
  url    = {https://huggingface.co/abhishekgautamm/lipi}
}