ByT5 multilingual P2G (small) — 300M params

Phoneme-to-grapheme inversion of the harmonized retrain. Input: <lang>: phoneme tokens. Output: word spelling. Trained on a 4.12M-pair harmonized corpus (same as the G2P sibling).

Results (4k stratified test sample)

	score
micro exact	0.483
macro exact	0.581

P2G is harder than G2P (phoneme → spelling is not one-to-one), but this model is useful for AAC scenarios where a user sounds out a word they can't spell.

Files

HF-format weights at root (~1.2 GB)
onnx/ — validated encoder+decoder pair

Licence

CC BY-SA 4.0

Downloads last month: 33

Safetensors

Model size

0.3B params

Tensor type

F32

Model tree for willwade/byt5-p2g-multilingual

Base model

google/byt5-small

Quantized

(6)

this model