gotham13's picture
Edge-Punct-Casing English from the sherpa-onnx package: model.int8.onnx, bpe.vocab, LICENSE and model card
a674d07 verified
|
Raw History Blame Contribute Delete
3.4 kB
metadata
license: apache-2.0
library_name: onnx
language:
  - en
base_model: frankyoujian/Edge-Punct-Casing
tags:
  - punctuation
  - truecasing
  - sherpa-onnx
  - onnx
inference: false

Edge-Punct-Casing (English), ONNX, for the browser

English punctuation and word casing for transcripts that have neither: a CNN-BiLSTM with a 5,000-piece unigram vocabulary by Jian You (Cisco Systems), from the paper "A lightweight and efficient punctuation and word casing prediction model for on-device streaming ASR" (code, weights). It adds commas, full stops and question marks after words and predicts lower case, capitalised or all-capital words.

This repository holds the two files of the sherpa-onnx package of that model, so that a web page can fetch them (the package is only published as a .tar.bz2 archive on GitHub releases, which a page cannot read). It is not an official release of the author or of sherpa-onnx.

What was changed

Nothing in the files. They were taken from sherpa-onnx-online-punct-en-2024-08-06.tar.bz2 (SHA-256 9f5e5a72c7d2829635bd074fce92b6bbd5b78da8a52e7ad8ed1be933f366b99d), which sherpa-onnx built from the author's sherpa-onnx-cnn-bilstm-unigram-bpe-en.7z. Only the 8-bit model is copied; the archive's float32 model.onnx (29.7 MB) is left out.

File Bytes SHA-256
model.int8.onnx 7,490,500 9d611f445fe4a46186080fe161be6059d87d72eb88d3a8cb00c1a06e83a6067e
bpe.vocab 149,430 e118b7ad88c54db562517df49e1cffd4836d166c34fb190fd311d7f34eb238f5

How it runs

The steps are sherpa-onnx's OnlinePunctuationCNNBiLSTMImpl and simple-sentencepiece:

  1. Split the text on ASCII white space. Cut each word, with ▁ in front, into vocabulary pieces: the cut with the highest sum of the pieces' scores (float32), the shorter first piece on a tie, <unk> (0) for a byte no piece starts with.
  2. Chunks of at most 200 tokens: <s> (1), the pieces, </s> (2), padded with 0. valid_ids is 1 for <s>, the first piece of each word and </s>, else 0; label_lens is the number of ones.
  3. token_ids and valid_ids int32 [N, 200], label_lens int32 [N] give active_case_logits and active_punct_logits, float32 [words, 4]: one row per word, chunks one after the other.
  4. Case (argmax): 0 as written, 1 all capitals, 2 first letter capital, 3 mixed (left as written). Punctuation (argmax): 0 none, 1 ,, 2 ., 3 ?, written after the word. Only ASCII letters change case.

A word longer than one chunk crashes sherpa-onnx 2.0.1; keep words under 190 bytes.

bolo runs it on its server through sherpa-onnx (punct-en-cnn-bilstm) and in its browser demo with ONNX Runtime Web, with a TypeScript port of these steps that gives the same text as sherpa-onnx on its test transcripts.

Licence and attribution

Apache-2.0 (see LICENSE, the licence of Edge-Punct-Casing and of its weights on the Hugging Face Hub). Model and training by Jian You (Cisco Systems); ONNX packaging by the sherpa-onnx project (Apache-2.0).