The intention of this repo is to provide a C++ interface to infer the pronunciation of unknown words. It's meant to be used in Tihu project.
The main repo that the models are trained with is here. After the encoder and decoder models are trained, they must be converted to the script module. Encoder and Decoder script modules must be passed to the G2P model which is defined at convertible_model.py.
Also grapheme and phoneme tables must be provided as text file. The assets must be located in resources folder.
A trained and converted g2p model plus the corresponding graphemes and phonemes text files are provided.
The Pre-cxx11 ABI version of Libtorch for cpu must be downloaded from here. Extract it in a folder named torch_deps. The torch_deps/libtorch folder and it's contents is a must have to compile this project.
Also shared objects that are necessary for running the binaries are located in torch_deps/libtorch/lib.
A make file is provided to compile this project, but before running the make command, create obj and build directories.
NOTE
Don't forget to set LD_LIBRARY_PATH before running the executables in build:
export LD_LIBRARY_PATH=torch_deps/libtorch/lib