willwade/byt5-p2g-multilingual

ByT5 multilingual P2G (IPA → grapheme)

One byte-level model, 122 languages, IPA pronunciation → word (spelling). Fine-tuned from google/byt5-small (3 epochs, 2.67M pairs).

Input convention: <lang>: IPA tokens (<eng>: h ə l oʊ). Output: the spelled word (hello). Handles made-up pronunciations (e.g. <eng>: m uː → moo).

Results (test split, 4k sample)

	exact	notes
micro (all 122)	0.509	macro 0.647; TER 0.491
Italian	0.967	vs 0.229 for the per-language WFST tier
German	0.924	vs 0.495 WFST
Spanish	0.180 raw	near-zero char error: valid b/v homophone spellings (abalanzares→avalanzares); homophone-credited scoring pending
English	0.302	homophone-rich; deep orthography
Mandarin/Cantonese	0.000	logographic

P2G is one-to-many (homophones); exact-match understates quality where valid alternative spellings exist.

Use as the OOV tier

Built for floravox as tier 3 behind FST lexicons (exact) and per-language Phonetisaurus WFSTs — the model only handles true OOV, so greedy decoding latency is acceptable and cached in practice.

HF-format weights (config/tokenizer/safetensors) will accompany the next release; the initial ONNX export was faulty and has been removed pending re-export. Numbers above are from the HF checkpoints.

Status: v1 preview (evaluation only)

These are v1 results — the model card numbers stand, but v1 weights are not available (lost in an infrastructure mishap before publishing; see repo history). v2 is imminent: retrained on a dialect-keyed corpus (eng-US/UK, por-BR/PT, spa-ES/LatAm, ...), with an improved recipe, in small (300M) and tiny (17M) sizes, with validated ONNX exports. Weights + ONNX + usage guide will land here shortly.

Training data & licence

Corpus merged from WikiPron (CC BY-SA 4.0; 117 languages) and gruut lexicons (MIT; 13 languages). Weights are released CC BY-SA 4.0 under the conservative reading that models trained on share-alike data inherit it. Attribution: Wiktionary/WikiPron (CUNY-CL), gruut (rhasspy), Google byt5 base (Apache-2.0).

Training code: voicegarden-lexicons/scripts/train_byt5 —one GPU-day on an RTX 3090 (~$6 rented).