git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3452 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

Changes to generate-proxy-keywords.cc so that it can print 'kwid weight proxies' to each line, in addition to print out the FST.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3451 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: BABEL scripts: small configuration changes (make more sane the combination of minibatch-size and max-change)

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3450 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

* nnet-forward.cc - disable GPU memory caching
* cmvn-to-nnet.cc - add variance flooring
* paste-feats.cc - cosmetic change option description
* train_nnet.sh - import cmvn settings from pre-training
* pretrain_dbn.sh - allow generating input transform from a prototype



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3449 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: bugfix, lattice conversion tool kaldi->HTK; represent time internally as integer number of frames, avoid comparing floats for exact match.



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3448 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: bugfix, lattice conversion tool kaldi->HTK, fixing timing bug, state sequence in FST weight can also be empty.



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3447 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: updating lattice conversion tool producing HTK-SLF, making it more robust (adding explicit terminal super-state, handling accepting_states with FST weight, words on nodes or links). Big thanks to Korbinian for help.



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3446 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: Change matrix library from using std::exp to exp/expf, for speed (thanks to tvarog for pointing the issue out); fix to determinize-lattice-pruned-test.cc to work with current code; disallow certain constructors in util/kaldi-table.h that would not work correctly.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3445 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3443 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3442 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3441 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3440 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

Changes to lattice determinization: 1. changed deque to priority_queue in EpsilonClosure in determinize-lattice-pruned.cc, which should speed up the determinization process substantially; 2. added KALDI_VLOG(2) debug info to lattice-determinize-pruned.cc and lattice-determinize-pruned-parallel.cc

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3439 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3438 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3437 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3435 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3434 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3433 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3432 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk : lattice align tools will produce topologically sorted lattices (can be useful for 3rd party lattice conversion scripts)



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3431 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3430 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3429 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3428 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3427 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

(trunk/babel/s5b) Changing the --zero-if-disjoint flag to --drop-frames flag during the call of train_mmi_sgmm2.sh

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3426 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: for speaker-id setup, changing the scripts to compute deltas before, not after, sliding-window CMN.  Caution: this requires retraining the system from scratch if you trained models using the old features.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3425 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: Change to keyword search (e.g. as used in BABEL program) to be more exact, e.g. with regard to time boundaries; make wspecifiers to scp support permissive (,p) option (see changes to docs for explanation); modify lattice-to-nbest to support random generation from lattice; add some nnet code for perturbing egs (no scripts to demonstrate this yet); misc cosmetic changes.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3424 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3423 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3422 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

Fix to discriminative training (retroactive thing and nnet-normalize-stddev), seems working, learning rate needs further tuning

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3421 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3420 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3419 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: change to word-align-lattice.cc so that lattice-align-words will output a partial lattice if not all could be aligned.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3418 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3417 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3416 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk,nnet: updating training scripts (now using prototypes, initializing with 'nnet-initialize'; NN training target represented as 'Posterior'), preparing NNs to allow parallel nested networks.



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3415 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3414 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: small change to ivector-extractor-acc-stats.cc (make it less memory-hungry when forking to open tables)

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3413 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3410 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk,nnet:minor changes to nnet components (<AddBias> <Rescale> are trainable, <Copy> adding initialization)



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3409 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8