punct_cap_seg_en — CoreML (int8) for PUSHTYPE

A CoreML conversion of 1-800-BAD-CODE/punctuation_fullstop_truecase_english (Apache 2.0), used by the PUSHTYPE dictation app to place commas, periods and question marks after on-device speech recognition. All credit for the model goes to its original author; this repository only repackages it.

PunctCapSeg.mlmodelc/ — compiled CoreML model (macOS 15+): int8 weights (per-block 32), FP32 activations.
vocab.tsv — the model's SentencePiece unigram vocabulary (spe_32k_lc_en.model), one piece<TAB>score<TAB>type per line.
LICENSE — Apache License 2.0.

Conversion

From the original at commit b26fd1c40e88678859048898218ea4edcc24c84a (punct_cap_seg_en.onnx): ONNX → PyTorch (onnx2torch) → coremltools 9.0, FP32 compute precision, then int8 linear weight quantization. Output matched the original ONNX model on all 15 test sentences; about 9 ms per sentence on Apple Silicon.

FP16 activations are avoided on purpose: the graph's attention-mask constant overflows in half precision and quality degrades silently.

Input and output

input_ids: int32 [1, 256] — BOS (1), the lowercased text's token ids, EOS (2), padded with 3.
post_preds [1, 256]: after each token, null / acronym / . / , / ?.
cap_preds [1, 256, 16]: per-character capitalization.
seg_preds, pre_preds: sentence boundaries, and an unused pre-label.

Changes from the original

Format only (ONNX to CoreML, weights quantized to int8) and the vocabulary exported to text. No retraining.

License

Apache 2.0, as the original.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for PUSHTYPE/punct-cap-seg-en-coreml

Base model

1-800-BAD-CODE/punctuation_fullstop_truecase_english

Finetuned

(7)

this model