kaldi-nn
Extended speech recognition neural network based on Kaldi for reproducible research
Changelog
Add ReLU, SoftHinge to nnet1
Add Pnorm, Maxout to nnet1
Optimize Maxout GPU kernel code (as fast as Pnorm)
Add LSTM w/o projection layer
Add egs/hkust for reproducing the results
Add ReLU result
Add Convolution and Maxpooling components, and test code
Add example for training ConvNet in nnet2
Add LSTM projected component in nnet2
Add per-utterance training binary
Add per-utterance options for egs-operation binaries