A language modeling toolkit built on top of PyTorch, embracing Kaldi-style single-purpose executables. Currently, it's still built for PyTorch 0.3.
This code (dozens of refactoring commits ago) has been used for experiments reported in an Interspeech 2018 submission (link to a PDF will be added soon).
For examples on training an i-vector FF LM, refer to folder exp/is2018-smmlm.
You'll need the implementation of SMM around on your system, kindly change balls/smm_itf/smm_ivec_extractor.py:10 to fit your setup.
- Further cleanup
- Update to current PyTorch
- Additional decoder structures (NCE, hierarchical softmax, etc.)