Modelpublic

klebster/g2p_multilingual_byT5_tiny_onnx

sourceHugging Facecc-by-4.0updated 4mo agoView on Hugging Face

0likes32downloads

Model Card

Multilingual G2P ByT5 Tiny — ONNX

ONNX export of charsiu/g2p_multilingual_byT5_tiny_16_layers_100. Converts written words to IPA transcriptions across 100 languages.


Architecture	ByT5-tiny (T5ForConditionalGeneration), 20.6M params
PER / WER	8.1% / 25.3% (100 langs, 500 words each, greedy)
ONNX FP32 size	106 MB (3 graphs: encoder + decoder + decoderwithpast)
ONNX INT8 size	27 MB (74% reduction, +0.0005 PER)
Best latency	5.30 ms/word (INT8, threads=8) — 1.5x faster than PyTorch CPU
License	CC BY 4.0

Quick Start

python

import onnxruntime as ort
from optimum.onnxruntime import ORTModelForSeq2SeqLM
from transformers import AutoTokenizer

so = ort.SessionOptions()
so.intra_op_num_threads = 8   # ~cores/4
so.inter_op_num_threads = 1

model = ORTModelForSeq2SeqLM.from_pretrained(
    "klebster/g2p_multilingual_byT5_tiny_onnx",
    provider="CPUExecutionProvider",
    session_options=so,
)
tokenizer = AutoTokenizer.from_pretrained("klebster/g2p_multilingual_byT5_tiny_onnx")

inputs = tokenizer("<eng-us>: hello", padding=True, add_special_tokens=False, return_tensors="pt")
preds = model.generate(**inputs, num_beams=1, max_length=50)
print(tokenizer.decode(preds[0], skip_special_tokens=True))
# Output: ˈhɛɫoʊ

Input format: <language_code>: word (e.g. <fra>: bonjour, <ger>: Straße). See CharsiuG2P for all 100 language codes.

Benchmark Summary

Tested on 10 languages, 300 words, greedy decoding. Hardware: Intel i9-13900KS, 128 GB DDR5.

Configuration	ms/word	vs PyTorch CPU
ONNX INT8 + threads=8	5.30	1.48x faster
ONNX FP32 + threads=8	7.47	1.05x faster
PyTorch CPU (baseline)	7.83	1.00x
ONNX default settings	25.09	0.31x

Key findings: Thread tuning is critical (default ONNX is 3x slower than PyTorch). INT8 quantization + threads=8 is the best config. GPU provides no benefit for this model size. Beam search provides negligible improvement (PER 0.0812 → 0.0808) at 3–8x latency cost.

Correctness: ONNX FP32 output is bit-identical to PyTorch. INT8 degrades PER by only 0.0005.

Evaluation

100 languages, 500 words each (50,000 total), CharsiuG2P test set:

Metric	PyTorch CPU	ONNX FP32	ONNX INT8
PER	0.0812	0.0812	0.0817
WER	0.2529	0.2529	0.2537

<details> <summary>PER by language (click to expand)</summary>

Language ISO-639-3 + Location / Script	PyTorch CPU	ONNX FP32	ONNX INT8
ady	0.0521	0.0521	0.0518
afr	0.0251	0.0251	0.0243
amh	0.3347	0.3347	0.3356
ang	0.0559	0.0559	0.0559
ara	0.6710	0.6710	0.6716
arg	0.0181	0.0181	0.0183
arm-e	0.0215	0.0215	0.0220
arm-w	0.0075	0.0075	0.0070
aze	0.0008	0.0008	0.0008
bak	0.0124	0.0124	0.0138
bel	0.0048	0.0048	0.0040
bos	0.0184	0.0184	0.0171
bul	0.0158	0.0158	0.0161
bur	0.1718	0.1718	0.1712
cat	0.1566	0.1566	0.1598
cze	0.0077	0.0077	0.0079
dan	0.2118	0.2118	0.2157
dut	0.0347	0.0347	0.0345
egy	0.3752	0.3752	0.3781
eng-uk	0.1542	0.1542	0.1534
eng-us	0.2182	0.2182	0.2193
enm	0.1988	0.1988	0.1979
epo	0.0002	0.0002	0.0000
est	0.0072	0.0072	0.0070
eus	0.0056	0.0056	0.0058
fas	0.1503	0.1503	0.1507
fin	0.0011	0.0011	0.0011
fra	0.0075	0.0075	0.0075
fra-qu	0.0041	0.0041	0.0041
geo	0.0254	0.0254	0.0258
ger	0.0454	0.0454	0.0442
gle	0.1775	0.1775	0.1802
glg	0.0690	0.0690	0.0687
grc	0.2944	0.2944	0.2942
gre	0.0232	0.0232	0.0229
hbs-cyrl	0.0967	0.0967	0.0972
hbs-latn	0.0900	0.0900	0.0913
hin	0.0421	0.0421	0.0417
hun	0.0227	0.0227	0.0224
ice	0.0281	0.0281	0.0273
ido	0.0484	0.0484	0.0479
ina	0.0529	0.0529	0.0532
ind	0.0219	0.0219	0.0231
isl	0.0341	0.0341	0.0341
ita	0.0273	0.0273	0.0282
jpn	0.1050	0.1050	0.1062
kaz	0.0055	0.0055	0.0061
khm	0.2787	0.2787	0.2850
kor	0.0776	0.0776	0.0776
kur	0.0093	0.0093	0.0099
lat-clas	0.0094	0.0094	0.0092
lat-eccl	0.0070	0.0070	0.0074
lit	0.0316	0.0316	0.0316
ltz	0.1288	0.1288	0.1307
mac	0.0113	0.0113	0.0110
mlt	0.0226	0.0226	0.0228
mri	0.2214	0.2214	0.2198
msa	0.0008	0.0008	0.0008
nan	0.1069	0.1069	0.1082
nob	0.0874	0.0874	0.0888
ori	0.0028	0.0028	0.0026
pap	0.0019	0.0019	0.0019
pol	0.0057	0.0057	0.0057
por-bz	0.0866	0.0866	0.0875
por-po	0.0872	0.0872	0.0877
ron	0.0072	0.0072	0.0072
rus	0.0211	0.0211	0.0202
san	0.0762	0.0762	0.0755
slk	0.0289	0.0289	0.0283
slo	0.0415	0.0415	0.0418
slv	0.1829	0.1829	0.1838
sme	0.0207	0.0207	0.0214
snd	0.2636	0.2636	0.2685
spa	0.0017	0.0017	0.0019
spa-latin	0.0005	0.0005	0.0007
spa-me	0.0043	0.0043	0.0047
sqi	0.0148	0.0148	0.0192
srp	0.0620	0.0620	0.0638
swa	0.0025	0.0025	0.0025
swe	0.0084	0.0084	0.0079
syc	0.3765	0.3765	0.3776
tam	0.0099	0.0099	0.0112
tat	0.0026	0.0026	0.0033
tgl	0.1066	0.1066	0.1066
tha	0.0921	0.0921	0.0929
tts	0.0196	0.0196	0.0199
tuk	0.0119	0.0119	0.0123
tur	0.0006	0.0006	0.0004
uig	0.0335	0.0335	0.0340
ukr	0.0338	0.0338	0.0332
urd	0.2352	0.2352	0.2374
uzb	0.0043	0.0043	0.0045
vie-c	0.0144	0.0144	0.0144
vie-n	0.0075	0.0075	0.0076
vie-s	0.0112	0.0112	0.0107
wel-nw	0.0967	0.0967	0.0990
wel-sw	0.1291	0.1291	0.1321
yue	0.2238	0.2238	0.2242
zho-s	0.2986	0.2986	0.2992
zho-t	0.3459	0.3459	0.3478
AVERAGE (arithmetic mean)	0.0812	0.0812	0.0817

</details>

<details> <summary>WER by language (click to expand)</summary>

Language ISO-639-3 + Location / Script	PyTorch CPU	ONNX FP32	ONNX INT8
ady	0.2700	0.2700	0.2680
afr	0.1240	0.1240	0.1220
amh	0.9940	0.9940	0.9940
ang	0.3260	0.3260	0.3200
ara	1.0000	1.0000	1.0000
arg	0.0960	0.0960	0.1000
arm-e	0.1120	0.1120	0.1140
arm-w	0.0520	0.0520	0.0480
aze	0.0060	0.0060	0.0060
bak	0.0920	0.0920	0.1020
bel	0.0540	0.0540	0.0440
bos	0.0620	0.0620	0.0600
bul	0.1080	0.1080	0.1140
bur	0.5080	0.5080	0.5000
cat	0.6000	0.6000	0.5980
cze	0.0420	0.0420	0.0440
dan	0.6740	0.6740	0.6700
dut	0.1940	0.1940	0.1920
egy	0.5820	0.5820	0.5920
eng-uk	0.3180	0.3180	0.3260
eng-us	0.4840	0.4840	0.4860
enm	0.7160	0.7160	0.7120
epo	0.0020	0.0020	0.0000
est	0.0640	0.0640	0.0620
eus	0.0220	0.0220	0.0220
fas	0.5640	0.5640	0.5660
fin	0.0080	0.0080	0.0080
fra	0.0200	0.0200	0.0200
fra-qu	0.0200	0.0200	0.0200
geo	0.2360	0.2360	0.2400
ger	0.1720	0.1720	0.1700
gle	0.6660	0.6660	0.6720
glg	0.3580	0.3580	0.3560
grc	0.7340	0.7340	0.7360
gre	0.1620	0.1620	0.1600
hbs-cyrl	0.5180	0.5180	0.5200
hbs-latn	0.4620	0.4620	0.4680
hin	0.2140	0.2140	0.2060
hun	0.1140	0.1140	0.1140
ice	0.1700	0.1700	0.1680
ido	0.2300	0.2300	0.2260
ina	0.2040	0.2040	0.2020
ind	0.1020	0.1020	0.1040
isl	0.2260	0.2260	0.2260
ita	0.2000	0.2000	0.2040
jpn	0.2380	0.2380	0.2400
kaz	0.0420	0.0420	0.0420
khm	0.7320	0.7320	0.7220
kor	0.0940	0.0940	0.0940
kur	0.0520	0.0520	0.0480
lat-clas	0.0580	0.0580	0.0560
lat-eccl	0.0380	0.0380	0.0400
lit	0.1860	0.1860	0.1900
ltz	0.3800	0.3800	0.3840
mac	0.0880	0.0880	0.0860
mlt	0.0880	0.0880	0.0900
mri	0.6740	0.6740	0.6740
msa	0.0040	0.0040	0.0040
nan	0.4100	0.4100	0.4140
nob	0.3400	0.3400	0.3340
ori	0.0220	0.0220	0.0200
pap	0.0120	0.0120	0.0120
pol	0.0240	0.0240	0.0240
por-bz	0.4320	0.4320	0.4320
por-po	0.3920	0.3920	0.3980
ron	0.0500	0.0500	0.0500
rus	0.1260	0.1260	0.1220
san	0.4400	0.4400	0.4360
slk	0.1140	0.1140	0.1160
slo	0.2180	0.2180	0.2200
slv	0.5860	0.5860	0.5880
sme	0.1360	0.1360	0.1380
snd	0.6000	0.6000	0.6120
spa	0.0160	0.0160	0.0180
spa-latin	0.0040	0.0040	0.0060
spa-me	0.0380	0.0380	0.0400
sqi	0.0740	0.0740	0.0900
srp	0.4480	0.4480	0.4440
swa	0.0180	0.0180	0.0180
swe	0.0600	0.0600	0.0560
syc	0.8980	0.8980	0.9080
tam	0.0500	0.0500	0.0560
tat	0.0140	0.0140	0.0180
tgl	0.4460	0.4460	0.4500
tha	0.2140	0.2140	0.2160
tts	0.0780	0.0780	0.0760
tuk	0.1100	0.1100	0.1140
tur	0.0040	0.0040	0.0020
uig	0.1480	0.1480	0.1540
ukr	0.2240	0.2240	0.2180
urd	0.5860	0.5860	0.5900
uzb	0.0220	0.0220	0.0240
vie-c	0.0400	0.0400	0.0400
vie-n	0.0300	0.0300	0.0300
vie-s	0.0400	0.0400	0.0380
wel-nw	0.3660	0.3660	0.3720
wel-sw	0.4120	0.4120	0.4260
yue	0.4040	0.4040	0.4080
zho-s	0.5320	0.5320	0.5300
zho-t	0.5600	0.5600	0.5620
AVERAGE (arithmetic mean)	0.2529	0.2529	0.2537

</details>

Citation

If you use this ONNX model, please cite both the original paper and the ONNX export:

bibtex

@misc{zhu2022byt5modelmassivelymultilingual,
      title={ByT5 model for massively multilingual grapheme-to-phoneme conversion},
      author={Jian Zhu and Cong Zhang and David Jurgens},
      year={2022},
      eprint={2204.03067},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2204.03067},
}

bibtex

@misc{noel2025g2pmultilingualbyT5tinyonnx,
      title={Multilingual G2P ByT5 Tiny — ONNX export},
      author={Kleber Noel},
      year={2026},
      month={apr},
      note={Published 2026-04-07},
      url={https://huggingface.co/klebster/g2p_multilingual_byT5_tiny_onnx},
}

Known Issues

German IPA quality: non-standard dialect

The model does not reliably produce Standard German (Hochdeutsch). Observed issues include use of alveolar flap /ɾ/ where Standard German uses uvular fricative /ʁ/, among other systematic deviations. See CharsiuG2P issue #20.

Spanish dialect dictionaries: `spa` and `spa-me` are identical

The spa (European Spanish) and spa-me (Mexican Spanish) dictionaries are identical in the upstream CharsiuG2P repository. They should differ in the /s/–/θ/ distinction (ceceo/seseo). See CharsiuG2P issue #15.