Views
No views yet
| Property | Value |
|---|---|
| Architecture | T5ForConditionalGeneration (ByT5) |
| Parameters | 20.8M |
| Encoder layers | 12 |
| Decoder layers | 4 |
| d_model | 256 |
| d_ff | 1024 |
| Vocab size | 384 (byte-level) |
| Format | SafeTensors (float32) |
| Size | ~83 MB |
| Metric | Score |
|---|---|
| PER (Phoneme Error Rate) | 0.096 |
| WER (Word Error Rate) | 0.281 |
1import CharsiuG2PKit
2
3let g2p = try G2P(modelDirectory: modelURL)
4let ipa = g2p.convert("hello", language: "eng-us")
5// "ˈhɛɫoʊ"1import mlx.core as mx
2from mlx.utils import tree_unflatten
3from safetensors import safe_open
4
5tensors = {}
6with safe_open("model.safetensors", framework="numpy") as f:
7 for key in f.keys():
8 tensors[key] = mx.array(f.get_tensor(key))pytorch_model.bin using:uv run scripts/convert_weights.py --model charsiu/g2p_multilingual_byT5_tiny_16_layers_100encoder.embed_tokens.weight, decoder.embed_tokens.weight) are deduplicated — only shared.weight is kept.1@inproceedings{zhu2022byt5,
2 title={ByT5 model for massively multilingual grapheme-to-phoneme conversion},
3 author={Zhu, Jian and Zhang, Cong and Jurgens, David},
4 booktitle={Interspeech},
5 year={2022}
6}