willwade/byt5-g2p-multilingual-tiny

ByT5 multilingual G2P (tiny) — 17M params

Byte-level seq2seq, 136 variety-keyed languages (eng-US/eng-UK, por-BR/por-PT, spa-ES/spa-LatAm, Welsh N/S, Armenian E/W, Bengali varieties, 20+ Sinitic splits). Trained on a 3.02M-pair corpus (WikiPron CC BY-SA + gruut MIT, dialect splits preserved) with language-balanced sampling.

Input: <lang>: word (ISO-639-3, variety-suffixed where split — <eng-US>: hello, <spa-ES>: abeja). Output: space-separated IPA in gruut/WikiPron conventions (h ɛ l ˈoʊ) — what piper-family voices consume directly.

Results (4k stratified test sample)

	score
micro exact	0.729
macro exact	0.654
token error rate	0.139

The tiny (17M) matches the small (300M) within noise on this task at 1/18th the size — see the sibling repo. Model artifacts: ~~70 MB.

Files

• HF-format weights at root
• onnx/ — validated encoder+decoder pair. Consume with onnx_reference.py (a minimal correct consumer). CRITICAL conventions: token id = byte + 3; EOS appended to encoder input; decoder needs an explicit causal mask and a length-2 bootstrap — the reference script encodes all of them.

Use as a tier

Built for floravox as the neural OOV tier behind FST lexicons and per-language Phonetisaurus WFSTs; cached in practice.

Licence

CC BY-SA 4.0 (share-alike inherited from WikiPron training data). Attribution: Wiktionary/WikiPron (CUNY-CL), gruut (rhasspy), Google byt5 base (Apache-2.0). Training code: voicegarden-lexicons/scripts/train_byt5.