scribe-punct-en
ONNX int8 punctuation + truecasing model for English, packaged for
Scribe offline voice dictation.
Restores . , ? ! : ; ' - and word capitalization on lowercase ASR output.
Source
Exported from felflare/bert-restore-punctuation
(MIT, by Daulet Nurmanbetov) — bert-base-uncased fine-tuned on English reviews.
Original F1 ~90%. Conversion: torch.onnx.export (opset 14) + ONNX Runtime
dynamic int8 quantization. No accuracy-sensitive changes, verified output parity
on sample dictation before upload.
Files
| File | Size | Purpose |
|---|---|---|
model.int8.onnx |
~107 MB | Quantized model (CPU, ~10 ms per phrase) |
tokenizer.json |
~0.7 MB | Fast WordPiece tokenizer |
model.json |
1 KB | Scribe sidecar: label map (OU/OO/.O/!O/,O/.U/...) and decoding style |
Label scheme
One label per word: <punctuation><case>, e.g. .U = period after this word +
capitalize this word; OO = nothing. See model.json.
Usage in Scribe
Placed under models/punct/en/ by Scribe's model manager. Works fully offline
via ONNX Runtime. No Python torch/transformers needed at runtime.
License
MIT (same as the source model). Attribution: felflare/bert-restore-punctuation.