RnnLMG2P | Grapheme-to-Phoneme conversion with Joint-Sequence RnnLMs

RnnLMG2P is a C++ library. RnnLMG2P has no bugs, it has no vulnerabilities and it has low support. You can download it from GitHub.

Hacking about with joint sequence RnnLMs. First reasonably fast version. Direct decoding with a joint-sequence RnnLM. This uses a representation of the RnnLM that is a bit more efficient than the default for the purposes of decoding. Synapse arrays are shared. The size of the input array is the same size as the hidden-layer. The output layer is not stored anywhere, but rather allocated when ComputeNet is called. The current copy of the hidden layer activations and associated word-history is stored in the search tokens. The BRnnLM.h is the beginning of a full reimplementation of RnnLM with a mind to customizing it for G2P. Main features currently under development: * Phoneme category features for histories * Future-features for graphemes * Generic class definitions * Full Bi-directional implementation. This will be folded back into Phonetisaurus after I make a bit more progress. This also addresses several issues with the default RnnLM where G2P is concerned: * Shuffling of training data * BPTT issues related to -independent option * Model compression. The neurons and much of the direct-connections synapse table can probably be compressed a good deal. These are mostly empty.