torch-ctcdecode

torch-ctcdecode is an implementation of Connectionist Temporal Classification (CTC) decoding for Torch. C++ code is borrowed liberally from TensorFlow.

Requirements

• C++11 compiler (tested with gcc 4.8.4 and Apple LLVM version 8.0.0)
• CMake 3.1 or later
• Torch7

CMake3 Installation

Ubuntu

$ sudo apt-get install cmake3

OSX

$ brew install cmake

Installation

$ git clone https://github.com/willfrey/torch-ctcdecode
$ cd torch-ctcdecode
$ luarocks make

Note About Third Party Libraries

The underlying C++ library utilizes a number of third party projects. These projects are downloaded, built, linked, and installed (if necessary) by CMake. Libraries are statically linked unless they are LGPL, in which case they are dynamically linked.

CMake will download the following libraries:

• Eigen3
• GFlags
• GLog
• KenLM (LGPL)

See the lib/ctcdecode/cmake/ directory for more details.

Using torch-ctcdecode

local ctcdecode = require 'ctcdecode'

See tests/test.lua for some examples. More documentation to come.