add blank_penalty for offline transducer

[chiiyeh]

Fix #541
Tested with sherpa-onnx-offline. High penalty will force out more words.

[csukuangfj]

Thanks! Left some minor comments. Otherwise, it looks good to me.

[csukuangfj]

Shall we also consider the case when batch_size > 1 ?

p_logit[0] is only for the first utterance.

We need to process

p_logit[vocab_size*i + 0]
// for in range(n)

You can move this if statement into the for statement below.

[chiiyeh]

Fixed!

[csukuangfj]

Could you also update

sherpa-onnx/sherpa-onnx/python/sherpa_onnx/offline_recognizer.py

Line 38 in b929124

def from_transducer(

to add an extra argument

blank_penalty: float = 0

and also add docstring for it?

Note that you need to pass the argument to

sherpa-onnx/sherpa-onnx/python/sherpa_onnx/offline_recognizer.py

Line 114 in b929124

recognizer_config = OfflineRecognizerConfig(

[chiiyeh]

Fixed!
I also noticed there is no docstring for the hotwords params, as well as the max_active_paths was not propagated into the OfflineRecognizerConfig. Should I raise two seperate PR to fix these?

[csukuangfj]

Should I raise two seperate PR to fix these?

Yes, please do that in a separate PR. Thanks!

[csukuangfj]

By the way, please merge the master branch into your current branch so the CI can pass.

[csukuangfj]

Could you fix the c++ code style check?
https://github.com/k2-fsa/sherpa-onnx/actions/runs/7648706009/job/20842763975#step:4:38

Check last commit
sherpa-onnx/csrc/math.h:100:  Lines should be <= 80 characters long  [whitespace/line_length] [2]
Done processing sherpa-onnx/csrc/math.h
Total errors found: 1
[FAILED] sherpa-onnx/csrc/math.h
Error: Process completed with exit code 1

You can run

pip install clang-format
cd /path/to/sherpa-onnx
./scripts/check_style_cpplint.sh
./scripts/check_style_cpplint.sh 1
./scripts/check_style_cpplint.sh 2

to check the style locally.

[csukuangfj]

Thank you for your first-time contribution!