zctc is a CTC beam search decoding library written in C++ and integrated into Python using pybind. This library provides efficient and accurate decoding for Connectionist Temporal Classification (CTC) based models.
- High-performance CTC beam search decoding
- Seamless integration with Python via pybind
- Easy to use API
- Currently
softmaxinputs only are supported, have to provide support forlog_softmaxandunnormalizedinputs too. - Currently this package only accepts BPE tokenized vocabulary, still have to add Character vocab and any other as per requirements.
- Should include tests to ensure seamless working of the CTC logic.
This package is still under development, and this is still a beta version. So, this package is not yet published in pypi.
Please refer this file for usage example.
In that file, we've compared the time stats between parlance version of CTCBeamDecoding and ours. With the involvement of both KenLM and Lexicon, we found ours nearly 3 to 4 times faster for longer sequence inputs. We've also made few other changes from their version of decoder, like, we operate on numpy arrays instead of torch tensors to make the torch dependency loosely bound, etc...
Contributions are welcome! Please open an issue or submit a pull request on GitHub.
- CTC for the original algorithm
For any questions or inquiries, please contact [durgaivelselvan.mn@zohocorp.com].