DeepPhonemizer en_us_cmudict_ipa_forward (safetensors)
US-English grapheme-to-phoneme model: spelling in, IPA out. It is
DeepPhonemizer's en_us_cmudict_ipa_forward
checkpoint (a 12.7M-parameter forward transformer trained with CTC on
cmudict-ipa), converted to safetensors so it can be
loaded without unpickling code. The original download link no longer works.
Used by speakcheck to pronounce words that are not in the CMU Pronouncing Dictionary (names, brands, new words).
Files
model.safetensors: the weights, in float16. Predictions are identical to the original float32 checkpoint on a 3,000-word test. The fixed sin/cos positional table is left out (rebuilt at load time), as is the word list embedded in the original checkpoint.config.json: model size and the input/output symbol tables.
The conversion script is
scripts/convert_deep_phonemizer.py,
and the inference code is
_deep_phonemizer.py.
Limits
- US English only; no stress marks.
- Very long words (over ~15 letters) can come out wrong.
- Acronyms are not handled by the model; speakcheck spells them out before calling it.
License
MIT, Copyright (c) 2021 Axel Springer News Media & Tech GmbH & Co. KG - Ideas Engineering. See LICENSE. Only the file format changed (safetensors, float16); the model is not retrained.
- Downloads last month
- 17