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 point that's not stressed enough is the massive amounts of good data that is needed to fuel these kind of systems. Data collection, cleaning, aggregation and other processes itself are humungous tasks.
