edge-punct-casing-en-onnx

edge-punct-casing-en-onnx

Unmodified mirror of the sherpa-onnx packaging of Edge-Punct-Casing, Jian You (Cisco)'s English punctuation + casing model, for use by Sokuji's local sentence-segmentation stage.

Upstream

• Code and weights: frankyoujian/Edge-Punct-Casing — Apache-2.0 (LICENSE in this repo is the verbatim upstream text).
• Packaging: k2-fsa/sherpa-onnx, sherpa-onnx-online-punct-en-2024-08-06 (int8 build), itself downloaded from frankyoujian/Edge-Punct-Casing on the Hugging Face Hub (sherpa-onnx-cnn-bilstm-unigram-bpe-en.7z).
• Architecture: a CNN-BiLSTM predicting punctuation (NO_PUNCT/COMMA/PERIOD/QUESTION) and casing (LOWER/UPPER/CAP/MIX_CASE) per whitespace-separated English word, encoded with a simple-sentencepiece v0.7 unigram tokenizer.

No conversion was performed. Both files here are byte-identical to the files sherpa-onnx ships under sherpa-onnx-online-punct-en-2024-08-06/ — this is already the smallest (int8) build sherpa-onnx offers, and it needed no requantization to reach parity with its own reference implementation (see "Measured numbers" below).

Files

bpe.vocab holds 5,000 piece<TAB>score lines with CRLF line endings.

Measured numbers

From docs/superpowers/notes/2026-09-14-asr-punctuation-benchmark.md (Sokuji repo, 2026-09-14):

• Parity vs. sherpa-onnx 1.13.8's own Python reference (OnlinePunctuation.add_punctuation_with_case), 97 rows covering 72 corpus rows plus 13 edge cases and 12 batched concatenations: 96/97 exact. The one difference is a single word in a 5-row batch where onnxruntime-web and native onnxruntime disagree on a near-tie logit (0.018 apart).
• Quality (en, stripped input): breakpoint F1 92.3 offline, sentence-end F1 84.5, casing F1 84.0; streaming breakpoints P/R 90.8 / 94.5, streaming sentence ends P/R 76.2 / 88.9 (8-character right-context commit). Best-scoring English model measured in the benchmark; it never forces a final mark, relying on the VAD end to supply it.
• Renderer cost (Electron 40.8.5, DGX Spark GB10 aarch64, WASM, no GPU needed): 7.6 MB on disk, 240 English characters in 15 ms, 960 characters in 39 ms.

License

Apache-2.0, carried over from frankyoujian/Edge-Punct-Casing. See LICENSE in this repository (verbatim copy of the upstream LICENSE file).