git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4927 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4926 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: In speaker ID system replaced PLDA length-normalization with the simpler form of length-normalization as performed in the binary ivector-normalize-length. This results in a 1% improvement for PLDA both before and after adaptation in the SRE08 example.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4925 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4924 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4923 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4922 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4921 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4920 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk,nnet1 : adding support to --one-silence-class, however this is probably not the main problem in sMBR step, need to have a look at this...



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4919 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4918 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4917 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

(trunk/egs/wsj) The script cleanup/find_bad_utts.sh now provides the user with much more additional info that could be useful when figuring what is wrong + some minor fixes of the error analysis scripts

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4916 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4915 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: src/nnet2/nnet-compute-discriminative.cc: Fixed minor bug in computing objective function stats for printing (nan for mmi numerator loglike)

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4914 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4910 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4909 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: Adding option to IvectorExtractor which diagonalizes the quadratic term in the objective function. This has the effect of reordering the iVectors' dimensions from most important to least important, which may make them easier to view.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4908 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4907 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4906 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4905 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

- lstm : implemented gradient-clipping, extended gradient print to show stats of forward / backward pass.
- nnet-loss.cc : bugfix in Xent::Eval (avoiding log(0) in cross-entropy computation)
- nnet-pdf-priors.{h,cc},nnet-forward.cc : refactored, behavior stays the same
- train-transitions : updates transition probabilities in transition model (but there was no WER improvment)




git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4904 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4903 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4902 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4901 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4900 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: adding new option --one-silence-class to sMBR/MPE discriminative training code and scripts... seems to be useful in ASPIRE to prevent discriminatively trained nnets generating too many insertions; need to test in other scenarios.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4899 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4896 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: bug-fix to table reading code that meant that when ,p option was used with scp: input, objects were re-read unnecessarily; this caused slowness in extract-segments (thanks to Yenda for noticing the problem).

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4895 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4894 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4893 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4892 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4891 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4890 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: some minor script extensions relating to online-nnet2 (thanks to Vijay for some of it); cosmetic code change.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4889 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4888 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4887 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: Decoder fixes so that the --min-active option is respected more accurately (should help speed too); efficiency improvements and extra option in CMVN code; various cosmetic improvements.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4886 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: changes to the librispeech setup to improve compatibility with systems using older versions of Bash(e.g. OS X)

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4882 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4881 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

Adding scripts incorporating word-dependent silence probabilities in making lexicon fst. Fixed a few bugs in some nnet2 scripts

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4880 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8