From the GitHub page: "This project contains our high performance CPU and CUDA versions of the CTC loss, along with bindings for Torch. The library provides a simple C interface, so that it is easy to integrate into deep learning frameworks."
The hard part is not the technology, it's the large enough data sets. Perhaps finding good employees could also be a help
Github [from article]: https://github.com/baidu-research/warp-ctc