CHANGELOG

[0.2.31.3] - 2021-06-17

Refactored pkwrap/ library. More refactoring to come!
Egs creation for e2e-LFMMI
- This now possible without dumping the supervision objects.
- This is useful when trained with large amounts of data.

[0.2.30.0] - 2021-04-20

Merged NGD implementation in pytorch
- recipe available for minilibrispeech

[0.2.28.2] - 2021-04-19

pytorch 1.8.1 and CUDA 11.1 compatibility tested
ChainModel refactoring
- initialization and model loading are separate functions
train.py: conveniently outputs best WER

[0.2.27.3] - 2021-03-19

librispeech i-vector recipe added
issues with decoding ChainModel when using i-vectors fixed

[0.2.27.1] - 2021-02-11

Patch to enable compatibility with all recent Kaldi versions (since Apr 2020)

[0.2.27] - 2021-02-03

Fixes to minilibrispeech recipe
- typos in tdnnf.py fixed
- missing parameters in config file added
train.py: regression fix to handle e2e option correctly

[0.2.26] - 2021-02-02

Librispeech 100h e2e recipe added
- train.py supports normal and flatstart training now
- ChainE2EModel class added to chain.py
xent_regularize is now actually a tunable parameter. Earlier it was fixed to be 0.1
train_lfmmi_one_iter can now take an optimizer
TDNN-F's constrain_orthonormal now uses add_ instead of addmm_
- addmm_ was not providing accurate results
missing copyrights added