ByT5 multilingual G2P (tiny) — 17M params
Byte-level seq2seq, 142 language/ variety tags (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 4.12M-pair harmonized corpus, with IPA conventions unified per language via a learned M2M aligner.
Input: <lang>: word (ISO-639-3, variety-suffixed where split
— <eng-US>: hello, <spa-ES>: abeja). Output: space-separated IPA
in gruut-anchored convention (h ɛ l ˈoʊ) — what Piper-family
voices consume directly.
Results (4k stratified test sample, same test set)
| this model | previous (0.729 on old test) | |
|---|---|---|
| micro exact | 0.718 | 0.582 (same test, same eval) |
| macro exact | 0.619 | 0.644 |
The tiny (17M) stays within striking distance of the small (300M) at 1/18th the size. The new corpus (+27% more entries, harmonized conventions) gives the tiny model a +13.6 pt lift on the same test set (0.582 → 0.718).
Files
- HF-format weights at root (~70 MB)
onnx/— validated encoder+decoder pair (manual TorchScript export, 3/3 gate-pair validation passing). Consume withonnx_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.
Quickstart (ONNX, CPU)
# pip install onnxruntime numpy
from huggingface_hub import snapshot_download
import sys; sys.path.insert(0, "voicegarden-lexicons/scripts/train_byt5")
from onnx_reference import load, run
d = snapshot_download("willwade/byt5-g2p-multilingual-tiny", allow_patterns=["onnx/*"])
enc, dec = load(d + "/onnx")
print(run(enc, dec, "<eng-US>: floravox"))
Training
The checkpoint was retrained on a 4.12M-pair harmonized corpus built
from gruut + WikiPron sources with M2M-convention alignment. Training
details in the
voicegarden-lexicons
repository (scripts/train_byt5/). Run: RTX 3090, ~8h, $1.20.
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.