| --- |
| license: mit |
| library_name: deep-phonemizer |
| tags: |
| - grapheme-to-phoneme |
| - g2p |
| - cmudict |
| - arpabet |
| - deep-phonemizer |
| --- |
| |
| # DeepPhonemizer US-English CMUdict model |
|
|
| This repository preserves the US-English ARPABET forward checkpoint referenced |
| by the upstream [DeepPhonemizer](https://github.com/axelspringer/DeepPhonemizer) |
| project and used as the out-of-vocabulary fallback by |
| [openWakeWord](https://github.com/dscripka/openWakeWord). |
|
|
| The upstream download host became unavailable, so this copy is retained at an |
| independent, stable location for reproducible Tater wake-word training. |
|
|
| ## Checkpoint |
|
|
| | File | Size | SHA-256 | |
| | --- | ---: | --- | |
| | `en_us_cmudict_forward.pt` | 66,725,366 bytes | `c2e1fb223d7e027bf7b33052540c6f71d19db6d7fd87ab8671152b8b114501c2` | |
|
|
| The checksum was verified against two independently hosted public copies before |
| this archive was created. Consumers should verify this SHA-256 after every |
| download. |
|
|
| This is the CMUdict/ARPABET checkpoint. It is not the similarly named IPA |
| checkpoint. |
|
|
| ## Loading and security |
|
|
| This is a legacy PyTorch checkpoint containing pickled DeepPhonemizer tokenizer |
| and preprocessor objects. Only load it after verifying the checksum and only in |
| an environment where the checkpoint is trusted. PyTorch 2.6 and newer default |
| to restricted `weights_only` loading, so software using DeepPhonemizer 0.0.19 |
| may need an explicit trusted legacy-checkpoint compatibility path. |
|
|
| ## Attribution |
|
|
| DeepPhonemizer and the pretrained model were produced by Axel Springer News |
| Media & Tech GmbH & Co. KG – Ideas Engineering. The upstream MIT license is |
| included in this repository. |
|
|
