svogunas/g2p-lt-lexicon

Lithuanian Pronunciation Lexicon (G2P) — 233 k words

A clean, consistent Lithuanian grapheme→phoneme lexicon: 233,390 words in standard orthography paired with their phoneme sequence in the LIEPA-3 phon alphabet. To our knowledge this is the largest open Lithuanian pronunciation dictionary — a rare, high-value asset for TTS, forced alignment, ASR lexicons, and linguistic research.

It is the training data behind the G2P models svogunas/g2p-lt-byt5-base (flagship) and svogunas/g2p-lt-byt5-small.

Contents

file	rows	description
lexicon.tsv	233,390	the full lexicon, word \t phonemes (UTF-8, one header row)
train.csv	230,390	training split (lexicon minus the held-out test words)
test.csv	3,000	held-out evaluation split (words unseen in train)

Columns: word (standard Lithuanian spelling, lowercase) and phonemes (space-separated phoneme symbols). The train/test split is the exact split used to report our G2P models' test PER, so results are reproducible.

from datasets import load_dataset
ds = load_dataset("svogunas/g2p-lt-lexicon")
print(ds["train"][0])   # {'word': 'aarono', 'phonemes': 'a a r oo n oo'}

Phoneme notation (LIEPA-3 phon alphabet)

Space-separated phonemes. Conventions:

• Lowercase base symbols for the core sounds.
• ' after a consonant marks palatalization (Lithuanian minkštumas): v', tS', n'.
• Doubled symbols mark length: aa, oo, ii, ee, uu.
• Digraphs for diphthongs/affricates: ie, uo, ea; tS (č), dZ (dž), ts (c), dz.
• Capitalised variants encode stress / length distinctions (e.g. aA, Aa, Oo, Ea).

Full inventory (~90 symbols):

a i e s s' n' t' t u oo k m r' n r k' l' p j j' m' I v' p' ee d d' l ii S' E v w
g' g uu U aA o S b' tS' b A Aa Ee Uo Z' aa Oo Ii Ea ea ie z J ts' eA z' Z uo O iI
W uU f' dZ' f oO N iE eE N' Uu Ie R' R uO x h L' ts M' x' h' L M tS dZ dz' dz

How it was built

1. Extraction. The LIEPA-3 phonetic corpus (phon/) ships word + phoneme TextGrids. We aligned the words and phones tiers by time to recover each word's pronunciation.
2. EMO. Phoneme annotations from the VDU Lithuanian emotional-speech corpus (EMO) were extracted the same way and added.
3. Merge & clean. LIEPA + EMO were merged at the variant level into one consistent notation: 99.5 % mean cross-source agreement, the dominant pronunciation kept per word, a handful of minority variants and 2 garbage rows pruned. The result is the 233 k-word lexicon here.
4. Excluded by design: the medical corpus (a different, bare-SAMPA notation — would need re-extraction from TextGrids to harmonise) and dialectal data (phonetic, not orthographic).

Intended use

• Train a G2P model (e.g. ByT5 grapheme→phoneme — see the linked models).
• TTS front-end — text → pronunciation → speech.
• Forced alignment lexicon (with MFA or similar).
• ASR pronunciation lexicon / linguistic study of Lithuanian phonology.

Limitations

• One pronunciation per word (the dominant variant); homographs are not disambiguated.
• Standard Lithuanian only; dialectal pronunciations are out of scope.
• Stress/length is encoded per the source corpus conventions — verify against your phoneme set.

License & attribution

CC BY 4.0. Derived from LIEPA-3 and the VDU EMO corpus, both CC BY 4.0, distributed via CLARIN-LT. Please attribute:

Lithuanian pronunciation lexicon derived from LIEPA-3 and the VDU Lithuanian emotional-speech corpus (EMO), distributed via CLARIN-LT (CC BY 4.0).

Citation

@misc{g2p-lt-lexicon,
  title  = {Lithuanian Pronunciation Lexicon (G2P), 233k words},
  author = {Smaliukas, Arūnas},
  year   = {2026},
  note   = {Extracted from LIEPA-3 phon TextGrids + EMO, merged and cleaned (CC BY 4.0)},
  url    = {https://huggingface.co/datasets/svogunas/g2p-lt-lexicon}
}