byt5-g2p-multilingual-tiny-kd (comparison artifact)
Not the recommended tiny model β use byt5-g2p-multilingual-tiny instead. This repo is the knowledge-distilled variant, kept for the training-recipe record.
Why it exists
For the v3 corpus we tested whether KD (distilling from the 300M teacher) beats training a 17M tiny from scratch:
| model | micro exact | macro exact |
|---|---|---|
| tiny from scratch (recommended) | 69.3% | 62.8% |
| tiny KD from 300M teacher (this repo) | 66.2% | 56.7% |
For G2P, scratch wins by ~3 points. The opposite held for P2G, where the KD tiny beats scratch β see byt5-p2g-multilingual-tiny.
Contents
- HF-format weights at root (~70 MB)
onnx/β fp32 encoder+decoder pair (initializer-promoted, single-file)onnx-int8/β dynamic-QInt8 quantized pair (~4x smaller, decode-gated)
Data
v3 corpus: 4.31M pairs / 151 languages / 1,672 phone symbols, single convention per language. All training data permissively licensed (CC BY-SA / BSD / MIT); see the voicegarden-lexicons repo.
- Downloads last month
- 44
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support