| --- |
| license: mit |
| base_model: huper29/huper_recognizer |
| tags: |
| - coreml |
| - phoneme-recognition |
| - ctc |
| - wavlm |
| - pronunciation |
| language: |
| - en |
| --- |
| |
| # HuPER recognizer, Core ML (int8) |
|
|
| A Core ML build of [huper29/huper_recognizer](https://huggingface.co/huper29/huper_recognizer) |
| (WavLM-large with a CTC head over 39 ARPAbet phones plus DX), made for Accent Lab, a macOS app |
| for English pronunciation practice, to check each sound on the Mac. |
|
|
| - `HuPER.mlmodelc`: compiled Core ML model, macOS 15 or later. Weights are int8 (linear, |
| symmetric, per channel); the math runs in FP32. |
| - `vocab.json`, `added_tokens.json`: the output labels, unchanged from the original. |
|
|
| ## Input and output |
|
|
| - `audio`: Float32 `[1, samples]`, 16 kHz mono, 1 to 20 s, normalized to zero mean and unit |
| variance (as the original's `Wav2Vec2FeatureExtractor` with `do_normalize`). |
| - `logprobs`: Float32 `[1, frames, 46]`, log-softmax over the labels, one frame per 20 ms. |
|
|
| ## How close it is to the original |
|
|
| On 17 lines of real speech (845 phones), goodness-of-pronunciation scores differ from the |
| PyTorch model by 0.34 on average, and 4 of the 845 phones change sides of a −2.0 threshold. |
| The best label per frame matches 98 % of the time. |
|
|
| ## License |
|
|
| MIT, as the original model. All credit for the model goes to the HuPER authors. |
|
|
