g2p_multilingual_byT5_tiny_onnx

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.

Quick Start

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.

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:

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

Links

• Paper: Zhu et al. (2022), Interspeech
• Original repo: CharsiuG2P
• Base model: charsiu/g2p_multilingual_byT5_tiny_16_layers_100
• ONNX export by: klebster for phone-similarity

Citation

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

@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},
}

@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.