Support adding punctuations to the speech recogntion result

[csukuangfj]

Usage

Download a model

Please see
https://github.com/k2-fsa/sherpa-onnx/releases/tag/punctuation-models

mkdir /tmp
cd /tmp
wget https://github.com/k2-fsa/sherpa-onnx/releases/download/punctuation-models/sherpa-onnx-punct-ct-transformer-zh-en-vocab272727-2024-04-12.tar.bz2
tar xvf sherpa-onnx-punct-ct-transformer-zh-en-vocab272727-2024-04-12.tar.bz2
rm tar xvf sherpa-onnx-punct-ct-transformer-zh-en-vocab272727-2024-04-12.tar.bz2

Build sherpa-onnx from source

cd /tmp
git clone https://github.com/k2-fsa/sherpa-onnx
cd sherpa-onnx
mkdir build
cd build
cmake ..
make

Test case 1 (Chinese + English)

cd /tmp
cd sherpa-onnx/build
./bin/sherpa-onnx-offline-punctuation \
  --debug=1 \
  --ct-transformer=/tmp/sherpa-onnx-punct-ct-transformer-zh-en-vocab272727-2024-04-12/model.onnx 
  '这是一个测试你好吗How are you我很好thank you are you ok谢谢你'

The output is

Num threads: 1
Elapsed seconds: 0.003 s
Input text: 这是一个测试你好吗How are you我很好thank you are you ok谢谢你
Output text: 这是一个测试，你好吗？How are you？我很好？thank you，are you ok。谢谢你。

Test case 2 (Chinese only)

cd /tmp
cd sherpa-onnx/build
./bin/sherpa-onnx-offline-punctuation \
  --debug=1 \
  --ct-transformer=/tmp/sherpa-onnx-punct-ct-transformer-zh-en-vocab272727-2024-04-12/model.onnx 
  '我们都是木头人不会说话不会动'

The output is

Num threads: 1
Elapsed seconds: 0.002 s
Input text: 我们都是木头人不会说话不会动
Output text: 我们都是木头人，不会说话不会动。

Test case 3 (English only)

cd /tmp
cd sherpa-onnx/build
./bin/sherpa-onnx-offline-punctuation \
  --debug=1 \
  --ct-transformer=/tmp/sherpa-onnx-punct-ct-transformer-zh-en-vocab272727-2024-04-12/model.onnx 
  'The African blogosphere is rapidly expanding bringing more voices online in the form of commentaries opinions analyses rants and poetry'

The output is

Num threads: 1
Elapsed seconds: 0.002 s
Input text: The African blogosphere is rapidly expanding bringing more voices online in the form of commentaries opinions analyses rants and poetry
Output text: The African blogosphere is rapidly expanding，bringing more voices online in the form of commentaries，opinions，analyses，rants and poetry。

[mark95]

I'm having a problem understanding how to train this model because it is in Chinese. Is it possible to train that model for other languages? What format dataset do I need to train it? Is there a training script for that model?