svogunas/g2p-lt-byt5-base

g2p-lt-byt5-base — Lithuanian grapheme-to-phoneme (G2P)

A byte-level ByT5 model that converts Lithuanian words in standard orthography into their phoneme sequence. It is the pronunciation front-end for Lithuanian TTS (and useful for forced alignment, lexicon expansion, and linguistic tooling).

• —Input: one word, standard Lithuanian spelling, lowercase — e.g. labas
• —Output: space-separated phonemes in the LIEPA-3 phon alphabet — e.g. l a b a s
• —Architecture: `google/byt5-base` (582 M), fully fine-tuned. Byte-level ⇒ no tokenizer issues with Lithuanian letters or phoneme symbols.

This is the flagship (most accurate). A smaller, ~2× faster variant for edge/embedded use is `svogunas/g2p-lt-byt5-small`.

Results

Held-out test set of 3,000 words unseen in training (true generalisation), word-level PER (phoneme error rate) and exact-match word accuracy:

For comparison, the small variant scores 1.98 % / 88.3 %. A larger byt5-large was trained and did not improve over base (2.57–2.70 % PER) — capacity past byt5-base is wasted on this data, so base is the recommended model.

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tok = AutoTokenizer.from_pretrained("svogunas/g2p-lt-byt5-base")
model = AutoModelForSeq2SeqLM.from_pretrained("svogunas/g2p-lt-byt5-base")

def g2p(word: str) -> str:
    ids = tok(word.lower(), return_tensors="pt").input_ids
    out = model.generate(ids, max_length=64)          # greedy; add num_beams=5 for a tiny gain
    return tok.decode(out[0], skip_special_tokens=True)

print(g2p("labas"))            # l a b a s
print(g2p("abaravičiene"))     # a b a r a v' i tS' ie n' e

Feed one word at a time (the model is trained on isolated words). For a sentence, split on whitespace, look words up, and handle punctuation/numbers in your own front-end.

Phoneme notation (LIEPA-3 phon alphabet)

Phonemes are space-separated. 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) observed in the training lexicon:

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

Training data

• —LIEPA-3 word+phoneme TextGrids (phon/) — the bulk of the lexicon.
• —EMO (Lithuanian emotional speech) phoneme annotations.
• —Merged at the variant level into a 233 k-word pronunciation lexicon, one consistent notation, 99.5 % mean cross-source agreement, dominant variant kept; 3 k words held out for the test above.
• —Excluded by design: medical (different SAMPA notation — needs re-extraction) and dialectal (phonetic, not orthographic, transcripts).

Fine-tuned from google/byt5-base, 5 epochs, lr 5e-4, load_best_model_at_end on eval loss.

Limitations

• —Isolated-word model: no sentence-level context, sandhi, or homograph disambiguation.
• —Trained on standard Lithuanian; dialectal pronunciation is out of scope.
• —Stress placement follows the lexicon's conventions; verify against your TTS phoneme set.

License & attribution

Released under CC BY 4.0. Training data (LIEPA-3, EMO) is CC BY 4.0 from VDU / CLARIN-LT — please attribute them when you use this model:

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

Citation

@misc{g2p-lt-byt5-base,
  title  = {g2p-lt-byt5-base: Lithuanian grapheme-to-phoneme (ByT5)},
  author = {Smaliukas, Arūnas},
  year   = {2026},
  note   = {Fine-tuned from google/byt5-base on a LIEPA-3 + EMO pronunciation lexicon (CC BY 4.0)},
  url    = {https://huggingface.co/svogunas/g2p-lt-byt5-base}
}