willwade/byt5-g2p-multilingual

ByT5 multilingual G2P (small) — 300M 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 (+27% over the previous 3.02M), 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 (0.731 on old test)
micro exact	0.730	0.663 (same test, same eval)
macro exact	0.617	0.620

The new model is trained on a 30% larger, harmonized corpus where source conventions (gruut + WikiPron) are aligned per language with a many-to-many EM aligner. On the same held-out test set, micro exact climbs +6.7 pts (0.663 → 0.730). The old published 0.731 was on a different test set with mixed conventions — the new 0.730 is on a harder test (one convention per language, no ties to exploit).

Highlight wins: Spanish 0.720 → 0.976, French 0.694 → 0.917, Portuguese 0.000 → 0.570 (was WikiPron-only, now gruut-anchored), Dutch 0.014 → 0.188.

Files

• HF-format weights at root (~1.2 GB)
• onnx/ — validated encoder+decoder pair (manual TorchScript export, 3/3 gate-pair validation passing). 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.

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", 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.