Versions of Kaldi

Versioning scheme

During its lifetime, Kaldi has has three different versioning methods. Originally Kaldi was a subversion (svn)-based project, and was hosted on Sourceforge. Then Kaldi was moved to github, and for some time the only version-number available was the git hash of the commit.

In January 2017 we introduced a version number scheme. The first version of Kaldi was 5.0.0, in recognition of the fact that the project had already existed for quite a long time. The basic scheme is major/minor/patch, but the "patch" version number may also encompass features (usually back-compatible ones). The "patch number" automatically increases whenever a commit to Kaldi is merged on github.

We only intend to change the major or minor version number when making relatively larger changes, or non-back compatible changes.

We always plan to recommend that Kaldi users check out the latest version of 'master', since actively supporting multiple versions would increase our workload.

Versions (and changes)

This section lists the version numbers of Kaldi with the commit messages for each patch commit (by "patch commit" we mean a commit that does not increase the major or minor version number). Each time we add a new major/minor version number we will include a longer section explaining the changes involved.

Version 5.0

This is the first major/minor version number after introducing the versioning scheme. The latest revision of version 5.0 is saved as branch "5.0" on github.

Below are commits corresponding to minor version numbers 5.0.x.

5.0.0 c160a9883 2017-01-05 [build] Adding versioning mechanism to Kaldi. This is version 5.0.0 (#1306)

5.0.1 53e33dc4b 2017-01-06 [scripts] nnet3 scripts: minor bug fixes in error-handling code (#1321)

5.0.2 f15be6387 2017-01-06 [scripts] minor updates of scripts in nnet1 (#1318)

5.0.3 735b2b149 2017-01-06 [egs] minor fix in fisher_swbd/s5/local/chain/run_blstm_6h.sh (#1320)

5.0.4 a4b209159 2017-01-07 added reverberation based augmentation recipe for TDNN+LSTM model on AMI; a fix to aspire recipe (#1314)

5.0.5 5cdbd1879 2017-01-07 [build] Makefile changes: print time for tests; dash-compatibility change (#1322)

5.0.6 e09e92f87 2017-01-08 [egs] minor fixes for mandarin recipes (#1325)

5.0.7 722df9ae6 2017-01-09 [build] Fix dash problems [set shell to /bin/bash in Makefile]

5.0.8 f646952b3 2017-01-09 [build] src/base/get_version.sh: change regexp syntax to support older bash versions. (#1327)

5.0.9 5f15d253b 2017-01-10 [egs] cosmetic; small typo fix in sre08 i-vector DNN extraction (#1330)

5.0.10 69b616d4a 2017-01-10 [src,egs,scripts] Remove sinusoid detection code and old SGMM code/scripts (#1329)

5.0.11 cd0680270 2017-01-10 [src] nnet3: fixed bug (LstmNonlinearityComponent::ZeroStats() was not implemented) (#1302)

5.0.12 748a4a923 2017-01-11 [src] Restore init-ubm.cc to sgmm2bin (was deleted when sgmmbin deleted)

5.0.13 95f0feca1 2017-01-11 [src] nnet1: replacing 'bc' by 'awk' in steps/nnet/train_scheduler.sh (#1333)

5.0.14 728b303df 2017-01-11 [scrips] fix bug in lmrescore_rnnlm_lat.sh when using G.carpa (#1334)

5.0.15 42291a66f 2017-01-12 [doc,egs] Fix path from lm to lmbin in kaldi for dummies and various egs/*/*/path.sh

5.0.16 421485ee9 2017-01-12 [src] nnet3: Remove deprecated options which are no longer used (#1335)

5.0.17 847035e7b 2017-01-13 [egs] Remove some unused scripts

5.0.18 ea25438dc 2017-01-13 [build] Fix return code of tools/extras/install_sacc.sh (#1337)

5.0.19 e1e7bbf5c 2017-01-13 [scripts] nnet3: fix to xconfig parsing to enable e.g. "input@-1" in xconfig descriptors (#1338)

5.0.20 1dabfa54c 2017-01-14 [src] cosmetic changes to rnnlm-related code (#1283)

5.0.21 4d2c78e23 2017-01-16 Some bug fixes to I/O code for nnet3

5.0.22 6c0a012d5 2017-01-16 Remove checking code which would fail if code updated in middle of training run.

5.0.23 f7b2fe754 2017-01-16 Bug fix (resolve failing test)

5.0.24 390aff4e5 2017-01-18 [egs] Fix typo in egs/csj/s5/local/csj_run_rnnlm.sh (#1351)

5.0.25 e8e5928a6 2017-01-19 [src] cudamatrix: added {Cu,}VectorBase::SetRandUniform(), + optimize/fix cu-rand code (#1352)

5.0.26 d117c955d 2017-01-20 [src] Some style-related fixes (and fix compiler warnings) in src/lm/

5.0.27 b9c1e8bce 2017-01-20 [src] Modify some tests that were slow on travis, to make them faster. (#1347)

5.0.28 df730e8e6 2017-01-21 [src] Implement CuMatrixBase<Real>::CopyColsFromVec() (#1361)

5.0.29 04404176c 2017-01-23 [src,scripts]: Adding dropout schedule option to nnet3 (#1248)

5.0.30 61510ca0d 2017-01-22 nnet3/report : Modified directory specification options in generate_plots.py (#1368)

5.0.31 4507183f3 2017-01-24 [build] fixing patch for OpenFst's compat.h for Windows build (#1373)

5.0.32 99b7d964d 2017-01-24 [scripts] lexicon learning: update missing defaults and help message; other fixes (#1360)

5.0.33 5c205eae3 2017-01-24 [scripts] nnet3/report : Added support for parsing latest compute_{train,valid}*.log (#1371)

5.0.34 f42b3ced0 2017-01-24 [scripts] nnet1,make_denlats: prevent crash when utt-ids contain slashes (#1374)

5.0.35 82167f971 2017-01-24 [scripts] remove import statement that was creating problems in nnet3 scripts.

5.0.36 9208165a5 2017-01-25 [src] nnet3: Add the "per-frame" option to DropoutComponent (#1324)

5.0.37 e227eda38 2017-01-25 [doc] Documentation changes; add scripts to automatically add documentation of patch versions.

5.0.38 0c1517c37 2017-01-25 [egs] swbd/s5c, chain recipes: fix hardcoded directory name (#1377)

5.0.39 c96a0aa0f 2017-01-26 [build]: use BSD-compatible mktemp in get_version*.sh; thanks; @andrely.

5.0.40 6ec2eb003 2017-01-26 [build]: fix to previous commit regarding usage of mktemp.

5.0.41 3b7fd1f92 2017-01-26 [scripts] Set path in utils/ssh.pl just like in slurm.pl or queue.pl to avoid path problems (#1379)

5.0.42 193bb9214 2017-01-26 [build]: minor bug fix in maintenance/documentation script.

5.0.43 5af1983f1 2017-01-27 [scripts] Fix nnet3 training scripts for bug introduced in #1371

5.0.44 2f8c33d13 2017-01-27 [scripts,egs] make steps/dict/learn_lexicon.sh more robust RE empty g2p prons (#1378)

5.0.45 9d53e051b 2017-01-29 [scripts] nnet3 training: fix for bug introduced when dropout scripting support was merged.

5.0.46 da591795e 2017-01-30 [scripts] steps/dict/learn_lexicon.sh: minor fixes (#1385)

5.0.47 01d96bf10 2017-02-01 [scripts] Fix bug in dropout code found by Tanel Alumae

5.0.48 67cabd026 2017-02-01 [build]: resolving OpenFst compilation issue with gcc-6.x (#1392)

5.0.49 f1d7891c5 2017-02-01 [egs] Add new graphemic system for Gale Arabic, with newer nnet scripts (#1298)

5.0.50 21cfe99c5 2017-02-02 [build] Windows build: generate missing base/version.h; cosmetic changes (#1397)

5.0.51 cd97bd22a 2017-02-02 [build]: Enable cross compilation, including to android. (#726)

5.0.52 c747ed5d5 2017-02-06 [build] fixing issue introduced in the previous win commit (#1399)

5.0.53 21c803145 2017-02-07 [egs] Fix to HKUST nnet2/3 scripts. (#1401)

5.0.54 dc454cc0a 2017-02-08 [egs,scripts,src] Add BABEL s5d recipe; various associated fixes (#1356)

5.0.55 9a1244ffb 2017-02-08 [egs] removing empty files in BABEL recipe (#1406)

5.0.56 2fd8da973 2017-02-09 [scripts] Give prepare_lang the option to generate more phone disambig symbols (#1408)

5.0.57 bcc71b67d 2017-02-09 [egs] sprakbanken recipe: add nnet3 examples.

5.0.58 bd73932fb 2017-02-10 [src] Get rid of clang 3.9 warnings in table-reading code (#1414)

5.0.59 cc1d677f9 2017-02-10 [egs] egs/fisher_swbd/s5/local/online/run_nnet2_ms.sh, change default stage to 0 (#1416)

5.0.60 922620255 2017-02-11 [egs] Add example scripts for Frisian-Dutch language (FAME! corpus)

5.0.61 37b5352f5 2017-02-11 [src,doc] Documentation updates; fixes to comments.

5.0.62 68cee215c 2017-02-17 [src] Make various tests faster, especially without GPU (#1428)

Version 5.1

Some of the major changes introduced in version 5.1 are:

  • Kaldi now requires C++11 to compile, and we support only the latest version of OpenFst (1.6.0). (This simplifies Kaldi's code, and will later enable the threading code to be rewritten to use C++11's better and more portable mechanisms).
  • The way chunk size and feature context is handled in nnet3 is changed to allow variable chunk size and shorter context at utterance boundaries. See Context and chunk-size in the "nnet3" setup for more information.
  • A new decoding mechanism, Looped decoding, is introduced in nnet3; this allows faster and more-easily-online decoding for recurrent setups (but only unidirectionally-recurrent ones, like LSTMs but not BLSTMs).
  • Online decoding with nnet3 models is now rewritten; it's faster and it supports models like LSTMs.
  • The sequence-training scripts in nnet3 are refactored and are now simpler and use less disk space.
  • There are scripts for segmentation of long transcribed audio files.

The latest revision of version 5.1 is saved as branch "5.1" on github.

Below are commits corresponding to minor version numbers 5.1.x.

5.1.0 68cee215c 2017-02-17 [src] Make various tests faster, especially without GPU (#1428)

5.1.1 214551996 2017-02-17 [build] Updating version file-- this commit marks version 5.1.0

5.1.2 9b9b561e2 2017-02-17 Kaldi 5.1.0 (#1342)

5.1.3 6477decfe 2017-02-17 [build] Change travis build to use clang; remove -rdynamic from CXXFLAGS (it's in LDFLAGS)

5.1.4 23338bf51 2017-02-19 [scripts] fix syntax error in validate_lang.pl [thanks: daniel galvez]

5.1.5 7d538e2bd 2017-02-19 [build] Increase OpenFst version 1.6.0->1.6.1. (#1434)

5.1.6 84585876f 2017-02-20 [build] Use github not sourceforge for IRSTLM [sourceforge repo dead?] (#1435)

5.1.7 d218412ac 2017-02-20 [egs] clean up the HKUST scripts and add scoring filters (#1436)

5.1.8 7b862c725 2017-02-20 [src,doc] Update version documentation with version 5.1; makefile fix; add details to info string for nnet3 component.

5.1.9 483192ed4 2017-02-20 [build] update .gitignore: ignore openfst-1.6.1 (#1439)

5.1.10 8c77d2cc9 2017-02-21 [src,egs,scripts] Support frame-subsampling with non-chain nnet3 models (#1238)

5.1.11 2c3df88f0 2017-02-21 [build] cloning IRSTLM, use https to access github, not ssh (#1441)

5.1.12 9866a4c4d 2017-02-22 [build] Update Travis configuration to get builds to complete on time. (#1443)

5.1.13 351da2886 2017-02-22 [egs] Iban recipe: MacOSX compatibility fixes (#1448)

5.1.14 ae4a5ef62 2017-02-23 [scripts] Make it so i-vector ID is not required for steps/nnet3/decode.sh

5.1.15 70f536071 2017-02-23 [build] Stop env.sh from crashing when set -u is active (#1451)

5.1.16 dff2a3f2c 2017-02-24 upgrade the sox calls to use more compatible options (#1453)

5.1.17 37bebd153 2017-02-25 [egs] fix typo in egs/swbd/s5c/local/nnet3/run_ivector_common.sh (#1452)

5.1.18 4cda60423 2017-02-25 [scripts] xconfig: make scripts work when LDA-like preconditioning layer is not used (#1447)

5.1.19 089e59679 2017-02-25 [build] Update OpenFst minimum version check to 1.6 in tools/Makefile. (#1455)

5.1.20 344e1adc1 2017-02-25 [src] Fix bug in decodable-online-looped.cc (prevent crash in nnet3 online decoding).

5.1.21 25b1299b6 2017-02-25 [egs] fix MacOSX incompatibilities in calls of paste (#1457)

5.1.22 d60e3cc26 2017-02-27 [egs,scripts] Misc script fixes; refactor wsj/s5 examples; update tedlium/s5_r2 (#1456)

5.1.23 3f180124f 2017-02-27 [scripts,egs] Replace SGE-specific options to queue.pl (etc.) with generic options (#1461)

5.1.24 1db4bbad2 2017-02-27 [scripts] Fix to nnet2 script (won't change results) (#1462)

5.1.25 ee45757d4 2017-02-28 [scripts] nnet3: remove import from tdnn in the xconfig script since it no longer exists (#1464)

5.1.26 0c4441ed4 2017-02-28 [egs] Add online-decoding example in some Swbd chain scripts (#1466)

5.1.27 4a58ab98f 2017-02-28 [scripts] fix_data_dir.sh, prevent function returning false in case reco2file_and_channel does not exist (#1467)

5.1.28 f6b011f71 2017-03-01 [src] kwsbin/transcripts-to-fsts.cc, fix bug introduced in the last babel recipe commit (#1470)

5.1.29 d5a9d5c94 2017-03-02 [src] add CUDA kernel for backprop of NormalizeLayer (#1458)

5.1.30 ba77d9766 2017-03-03 [src] fix kws pipeline after upgrading to openfst-1.6 (#1471)

5.1.31 828f0b9cc 2017-03-03 [egs] Add discriminative training script for WSJ/TDNN-LSTM (#1468)

5.1.32 5bad9b2d9 2017-03-06 [src] bug-fix in gst plugin code (issue on g++ 5.4.0) (#1479)

5.1.33 34128c51a 2017-03-06 [scripts,egs] sMBR on LFR xent system with shifted feats (#1477)

5.1.34 5bfa9e541 2017-03-08 [egs] Minor cosmetic changes in voxforge example script (#1483)

5.1.35 bd23a10ed 2017-03-10 [egs] Fixes to URLs in vystadial example script.

5.1.36 1a4dbf6be 2017-03-13 [src] nnet1: fixing issue in multi-task training (#1491)

5.1.37 eba49a03f 2017-03-14 [build] Bump OpenFst version to v1.6.2 (#1492)

5.1.38 5c98096d8 2017-03-16 [egs] swbd/chain : added blstm script using fast-LSTM; added BLSTM+TDNN script. (#1497)

5.1.39 3fec956be 2017-03-17 [egs] update fisher_swbd recipe (fixes to how things are installed). (#1498)

5.1.40 eac0baa6b 2017-03-19 [src] sort cuda kernel function declarations to make searching easier. (#1501)

5.1.41 25ca8e4b0 2017-03-19 [build] Android compilation, bug-fixes (#1502)

5.1.42 ed33b1fd3 2017-03-19 [doc] Add a note to README.md about Android cross compilation (#1503)

5.1.43 965cc494c 2017-03-20 [egs] ami : Added tdnn_lstm recipe with fast-lstmp layer. Added tdnn_lstm recipe with -1 delay at lowest lstm layer (#1505)

5.1.44 39ad4103e 2017-03-22 [scripts] prevent failure when final.ie.id doesn't exist (#1508)

5.1.45 ef415a7a4 2017-03-22 [src] Fix exit code of extract-rows.cc (#1510)

5.1.46 9a6feea7b 2017-03-23 [egs] fixes to babel pipeline; thanks to Fred Richardson (#1509)

5.1.47 0d0316b48 2017-03-23 [src,scripts]: Several unrelated cosmetic changes

5.1.48 24bc2221f 2017-03-27 [misc] remove eXecute permissions where not needed (#1515)

5.1.49 244251fcc 2017-03-28 [egs] Fix to egs/wsj/s5/run.sh (unset variable) (#1517)

5.1.50 8b458ae1b 2017-03-29 [src] Adding noexcept to hashing function objects (#1519)

5.1.51 13d300fc6 2017-03-29 [src,doc] Fix several unrelated minor problems. Thanks: gaoxinglong

5.1.52 e9d79939b 2017-03-31 [src] (minor) Added missing SetZero() to NaturalGradientAffineComponent::Scale() if scale==0.0 (#1522)

5.1.53 e8ca375da 2017-04-01 [scripts,egs] Adding options for using PCA instead of LDA+MLLT for ivectors used in ASR. Results are reported in the default TDNN recipe in AMI. Updating steps/online/nnet2/{train_diag_ubm.sh,train_ivector_extractor.sh} so that they now backup the contents of their destination directory if it already exists. (#1514)

5.1.54 60f932721 2017-04-02 [build,src,doc] Modify get_version.sh to deal better with whitespace (avoid space in version); minor fixes (#1526)

5.1.55 3c9440103 2017-04-04 [build]: remove openfst check (#1531)

5.1.56 d8b34d4b5 2017-04-04 [src] cudamatrix: speed up AddColSumMat with transfrom reduce kernel template (#1530)

5.1.57 e5b141943 2017-04-04 [src] Cosmetic change: remove 'train.tra' from usage messages (#1529)

5.1.58 015768678 2017-04-05 [src] nnet1: improving the GPU diagnostics, (#1532)

5.1.59 d9efe6efa 2017-04-11 [src] Fix copy-feats for using the --write-num-frames and --compress true flags at the same time (#1541)

5.1.60 5b70933c6 2017-04-12 [scripts] fix to get_egs_targets.sh (thanks: David Pye)

5.1.61 b2d5d3683 2017-04-12 [scripts] nnet3 : fix issue where LDA estimation failed for LSTMs with label delay (#1540)

5.1.62 31a528b0c 2017-04-13 [src] Change ConvertStringToReal to be locale-independent (i.e. always-US). Fixes android issue. (#1513)

5.1.63 7fcc311c1 2017-04-13 [egs,scripts,src] SID and LID tools and scripts: cosmetic improvements, better error-handling, and various minor fixes; results unchanged. (#1543)

5.1.64 179d9b63c 2017-04-14 [src] add template instantiations for ConvertStringToReal, address issue #1544

5.1.65 889175079 2017-04-14 [egs] AMI TDNN Results Update (#1545)

5.1.66 9e06caa1f 2017-04-15 [srcipts] steps/nnet3/report/generate_plots.py: plot 5,50,95th percentile of value and derivative instead of mean+-stddev (#1472)

5.1.67 81346fc4b 2017-04-16 Update travis.yml so PRs to kaldi_52 are built

5.1.68 46f328527 2017-04-19 [egs] Added check for kaldi_lm being installed in fisher_swbd recipe. (#1558)

5.1.69 72234621b 2017-04-20 [doc] Small edit to hmm.dox, clarifying something

5.1.70 d82bbd7a2 2017-04-20 [src] nnet3 online silence weighting - adding frame subsampling factor (#1559)

5.1.71 d8be99a1c 2017-04-20 [src,scripts,egs] Add dropout for nnet3 LSTMs, with recipes. (#1537)

5.1.72 015c96f74 2017-04-20 [src] Add documentation for dropout function.

5.1.73 73489ae53 2017-04-21 [src] Fix to testing code signal-test.cc, change threshold to resolve failure (#1565)

5.1.74 87d95c5ef 2017-04-21 [egs] Add mini-librispeech example scripts [intended as a sanity-checker/tutorial setup] (#1566)

5.1.75 c68a576b0 2017-04-21 [egs] adding results and cleanup in mini-librispeech

5.1.76 1f51ef5d0 2017-04-23 [egs] Add soft link for mini-librispeech setup

5.1.77 06047be40 2017-04-25 [src] Minor bug-fixes in compute-wer-bootci and WSJ run.sh. Thanks: @osadjadi

5.1.78 8b9e8743c 2017-04-25 [egs] bug-fix in egs/ami/s5/run_ihm.sh (#1577)

5.1.79 a39e580db 2017-04-28 [build] Remove warning about dash (I believe we have no problem with it now)

5.1.80 fc20b7670 2017-05-01 [egs] fix small bug in wsj run.sh [thanks: Hang Lv]

5.1.81 e527572a3 2017-05-01 [sripts] extract_ivectors.sh: cosmetic fix, correct example usage. (#1600)

5.1.82 1cd6d2ac3 2017-05-02 [egs] mini-librispeech: download data from openslr if it's not present (#1598)

5.1.83 3906ae94a 2017-05-04 [src] bug-fix for memory bug in nnet computation cache code

5.1.84 ba138485f 2017-05-04 [src] Fix bug that can cause crash for for kernel-sizes=n*65536.

5.1.85 ad6491ab5 2017-05-05 [src] Add --write-per-frame-acoustic-loglikes options to alignment programs (#1601)

5.1.86 7c04d6291 2017-05-06 [scripts] nnet1: minor update i-vector and mpe scripts (#1607)

5.1.87 578c17622 2017-05-07 [src] cosmetic change to const-arpa-lm-building code; remove too-general template. (#1610)

5.1.88 6eebc0482 2017-05-09 [src,scripts,egs] Segmenting long erroneous recordings (#1167)

5.1.89 ab5c96c87 2017-05-11 [egs] thchs30 cmd and stage bug fix (#1619)

5.1.90 b1e860171 2017-05-13 [src] Change to GPU synchronization, for speed (disables GPU stats by default) (#1617)

5.1.91 8cbb74c43 2017-05-13 [src] Fix template instantiation bug causing failure if DOUBLEPRECISION=1

5.1.92 66666823a 2017-05-15 [egs,scripts] Updates to BUT-specific cmd.sh settings (affects only Brno team); changes RE verbose level in nnet1 scripts.

5.1.93 a7e6183c4 2017-05-15 [src] fix a small bug: logging cuda elapsed time (#1623)

5.1.94 7af2128d6 2017-05-17 [src,scripts,egs] Add capability for multilingual training with nnet3; babel_multilang example.

5.1.95 963895084 2017-05-18 [doc] small fix RE queue configuration.

5.1.96 ae45ee2ae 2017-05-19 [scripts] Fix to long-utterance segmentation script (#1631)

5.1.97 45ee31fb3 2017-05-20 [src] Make parsing error-msg more informative (thanks: Stefan-Adrian Toma)

5.1.98 b13f56bd2 2017-05-21 [src] fix regarding first/last chunk's right-context in chain models (#1632)

5.1.99 86766bdad 2017-05-21 [scripts, egs]: fix to egs/lre07/v2 (test was trained on); other updates to LRE scripts.

5.1.100 dc32e0b5e 2017-05-22 [scripts] Fix steps/cleanup/make_biased_lm_graphs.sh to actually add the top-n-words into the lms (#1637)

5.1.101 d6cf1bde9 2017-05-23 [scripts,egs] Fixes to long-recording segmentation (#1639)

5.1.102 b5eac0982 2017-05-24 [scripts] fix bugs in align_basis_fmllr.sh [thanks: Filip Jurcicek]

5.1.103 0f5e16ed3 2017-05-25 [build] Check python version is 2.7*, not just 2.*.

5.1.104 31683a38e 2017-05-25 [build] Change check_dependencies.sh to not look for yum if apt-get present.

5.1.105 6739d6020 2017-05-25 [scripts] Make more informative error in validate_lang.pl when path.sh prints something

5.1.106 6cc8e3ad7 2017-05-26 [src] Use STL thread support library instead of pthread. (#1350)

5.1.107 c210b444d 2017-05-26 [scripts] Fix bugs in automatic report generation for nnet3 training

5.1.108 76c8c777f 2017-05-26 [egs] Ported Fisher spanish recipe to use new LDC dir structure. Other small fixes (#1647)

5.1.109 6b906a373 2017-05-26 [egs] updated the LDC web address for wsj0-train-spkrinfo.txt (#1648)

5.1.110 0b0dbb016 2017-05-26 [scripts,egs] minor script fix; fixes in various recipes (#1649)

5.1.111 b6c34db1a 2017-05-26 [scripts] Lexicon expansion script -- fix for LM-probs, make it work for non-ASCII langs or langs w. large grapheme set (#1650)

5.1.112 e9646f031 2017-05-26 [src] keyword search: fix invalid assumption about the end states (#1651)

5.1.113 93c069cec 2017-05-27 [src] Fix bug in newly refactored threading code

5.1.114 726728154 2017-05-27 [scripts] Getting egs, limit max open filehandles to 512 (thanks: gaoxinglong9999)

Version 5.2

Some of the changes introduced between 5.1 and 5.2 are:

  • Upgrades to nnet3 to support batch-norm and convolutional components; recipes for certain image tasks (like CIFAR).
  • nnet3 training script simplifications and refactoring.
  • Some of the recipes are upgraded to include dropout and the –proportional-shrink option (which approximates l2 regularization); this improves results.

Many changes were made in the commits listed below (i.e. in the minor versions 5.2.x), including:

The latest revision of version 5.2 is saved as branch "5.2" on github.

Below are commits corresponding to minor version numbers 5.1.x.

5.2.0 726728154 2017-05-27 [scripts] Getting egs, limit max open filehandles to 512 (thanks: gaoxinglong9999)

5.2.1 393ef73ca 2017-05-29 [build] Upgrade .version (this is official start of kaldi 5.2)

5.2.2 cb7d17638 2017-05-29 Merge pull request #1547 from kaldi-asr/kaldi_52

5.2.3 a57eab677 2017-05-29 [scripts] Fix bug in PR #1646 (#1658)

5.2.4 972efd7a2 2017-05-31 [scripts] nnet3 script cleanups; add --proportional-shrink in more places. (#1659)

5.2.5 3e57783cf 2017-05-31 [src] minor bugfix in convolutional component (doesn't affect experiments)

5.2.6 44cd38652 2017-05-31 [build,scripts] Update scripts that make version info; remove no-op option from script.

5.2.7 fa198bb59 2017-06-01 [scripts] do not fail if the ivector extractor belongs to different user (#1662)

5.2.8 bd5fd04a1 2017-06-01 [src] keyword search tools: fix Minimize() call, necessary due to OpenFst upgrade (#1663)

5.2.9 ada5c0d80 2017-06-01 [src,scripts] Use ConstFst or decoding (half the memory; slightly faster). (#1661)

5.2.10 cdb25d114 2017-06-02 [scripts] make errors getting report non-fatal (thx: Miguel Jette); add comment RE dropout proportion

5.2.11 7728a10bc 2017-06-02 [src] Fix bug in online2-nnet3 decoding RE dropout+batch-norm (thanks: Wonkyum Lee)

5.2.12 6fe140add 2017-06-02 [src] Add #include missing from previous commit

5.2.13 13078031c 2017-06-07 [src] nnet3: fix small bug in checking code. Thanks: @maddin2000.

5.2.14 bf6ad5410 2017-06-07 [doc] Fix typo (#1669)

5.2.15 a0795ec52 2017-06-07 [scripts] Add python script to convert nnet2 to nnet3 models (#1611)

5.2.16 712c234db 2017-06-08 [scripts] Fix small bug in utils/gen_topo.pl.

5.2.17 5f680387d 2017-06-08 [build] Fix small bug in configure

5.2.18 c4dd63ac8 2017-06-08 [egs] Fix to some --proportional-shrink related example scripts (#1674)

5.2.19 7fa5f812d 2017-06-10 [src] Cosmetic fixes to usage messages

5.2.20 468c619a2 2017-06-13 [src] fix to gst-plugin/Makefile (remove -lkaldi-thread) (#1680)

5.2.21 defc8bd4e 2017-06-13 [build] when installing sequitur and mmseg, look for lib64 as well (thanks: @akshayc11) (#1677)

5.2.22 594daacc0 2017-06-13 [src] Fix bug in looped computation (#1673)

5.2.23 d46890dad 2017-06-13 [egs,scripts,src] Tune image recognition examples; minor small changes. (#1682)

5.2.24 08dbc187f 2017-06-14 [egs] CNN+TDNN+LSTM experiments on AMI (#1685)

5.2.25 ecc6a78a5 2017-06-14 [src,scripts,egs] Backstitch code+scripts, and one experiment, will add more later. (#1605)

5.2.26 9d494d250 2017-06-14 [egs] Fix failure in multilingual BABEL recipe (regenerate cmvn.scp) (#1686)

5.2.27 4d7ec0134 2017-06-15 [src] Fix to multiple-fst case of latgen-faster-mapped-parallel (memory bug) (#1688)

5.2.28 16b9918f3 2017-06-16 [egs] Rename files with Windows-incompatible names (#1690)

5.2.29 d76f088a0 2017-06-16 [src] nnet3 model-collapsing code, for slight decoding speedup (#1671)

5.2.30 20fe876fe 2017-06-19 [egs] babel recipe: check if icu4c is installed (#1697)

5.2.31 f7898b417 2017-06-20 [scripts] in subsegment_data_dir.sh, warn if utt2num_frames missing, etc. (#1702)

5.2.32 0a37d3e7e 2017-06-20 [src] Adding options to MBR/confidence code (#1696)

5.2.33 6a0768aa8 2017-06-20 [src] Fix compiler warnings and work around bug on Windows (#1698)

5.2.34 2ef9f9153 2017-06-21 [src] nnet3: fix assertion that shouldn't have been there. Thanks: @vimalmanohar

5.2.35 76c31c2e3 2017-06-21 [egs] fix problems in multilingual BABEL setup (#1691)

5.2.36 1022e9bd3 2017-06-22 [src] Fix bug in fstrmymbols RE recent const-fst changes (thanks: Jon Nichols); other cosmetic changes.

5.2.37 98d2fc7c2 2017-06-22 [egs] small update to librispeech recipe, RE const-FST.

5.2.38 a4578c4d6 2017-06-22 [egs] Adding hub4-ne broadcast spanish recipe (#1665)

5.2.39 b82b82f78 2017-06-23 [build] Update README.md / fix ci badge (#1709)

5.2.40 acb5439bf 2017-06-24 [egs,scripts]: replace non-portable read-link -f with utils/make_absolute.sh (#1694)

5.2.41 a75fb1be0 2017-06-26 [build] IRSTLM build: resolve problems with compilers by patching configure.ac (#1713)

5.2.42 456fec60f 2017-06-27 [egs] improve TDNN model in tedlium example (fewer jobs, proportinal-shrink 20) (#1715)

5.2.43 0d62760e6 2017-06-27 [build] update tools/extras/install_speex.sh to address #1718 (#1719)

5.2.44 6ae3111a1 2017-06-27 [scripts] Quote '{' in perl regexp (#1721)

5.2.45 913211dfd 2017-06-27 [scripts] Remove bogus note on CUDA non-use from compute_average_posterior() (#1722)

5.2.46 c12c1b807 2017-06-27 [scripts] run.pl: Default concurrent jobs to number of GPUs (#1723)

5.2.47 2d011df4a 2017-06-28 [src] Fix for threaded nnet2 decoding: check if threads are joinable before calling join(), to avoid multiple calls to join() (#1725)

5.2.48 f6f784bed 2017-06-28 [src] Fix bug in lattice string-pushing, affecting lattice-push (#1724)

5.2.49 3505e86be 2017-06-28 [scripts,egs] simplify nnet3 scripts by removing unused feature types (LDA, delta); add sMBR recipe for mini-librispeech (#1711)

5.2.50 16f39dedc 2017-06-29 [scripts] bug-fix in nnet3 scripts: change type of max_lda_jobs to int (#1728)

5.2.51 29c81102d 2017-06-30 [src] Bug-fix in model-collapsing code (thanks: aarora8)

5.2.52 2a0781897 2017-07-03 [src] Make sure softmax-related functions can work in-place. (#1729)

5.2.53 cac072dfa 2017-07-03 [egs] Add updated TDNN+BLSTM scripts for swbd/s5c, with dropout etc. (#1730)

5.2.54 abf898f42 2017-07-06 [src] remove remap-leaves.h (thanks: @kkm000)

5.2.55 28c613966 2017-07-07 [egs] APIAI example: model downloads links changed (#1747)

5.2.56 39c6ddede 2017-07-08 [egs] Update Librispeech nnet3 TDNN recipe (old one did not run) (#1727)

5.2.57 12ad1f63c 2017-07-09 [build] Add configuration check for incompatible g++ compilers when CUDA is enabled. (#1749)

5.2.58 97922a856 2017-07-09 [egs] Small fix to ami/s5b/local/chain/compare_wer_general.sh (#1751)

5.2.59 1644b5bb2 2017-07-11 [egs] some small updates to scripts (installing beamformit; segmentation example)

5.2.60 113e2a719 2017-07-11 [src] Fix usage message of program (thanks:jubang0219@gmail.com)

5.2.61 6aa5ed177 2017-07-11 [src] Fix indexing error in nnet1::Convolutional2DComponent (#1755)

5.2.62 b813b9d79 2017-07-11 [scripts] Fix bug in segment_long_utterances.sh (#1758)

5.2.63 a234281c7 2017-07-11 [egs] swbd/s5c, added 5 layer (b)lstm recipes (#1759)

5.2.64 40620fca8 2017-07-14 [src] Change name of option --norm-mean->--norm-means for consistency, thanks: 415198468@qq.com

5.2.65 2fbf2df59 2017-07-14 [src] Add support for row/column ranges when reading GeneralMatrix (#1761)

5.2.66 42c059625 2017-07-16 [scripts] minor bugfix to nnet1 alignment script when creating lattices (#1764)

5.2.67 b2fc6c4ee 2017-07-16 [scripts] For nnet3 training, add option to disable the model-combination (#1757)

5.2.68 6937b98cf 2017-07-16 [src] Change to arpa-reading code to accept blank lines with whitespace (#1752)

5.2.69 aedc2fe6f 2017-07-17 [egs] Add recipe for aishell data (free Mandarin corpus, 170 hours total) (#1742)

5.2.70 c794d5584 2017-07-19 [egs] Fixes to the aishell (Mandarin) recipe (#1770)

5.2.71 98bbe11b0 2017-07-19 [egs] Update AMI chain experiments RE dropout, decay-time and proportional-shrink (#1732)

5.2.72 47f8057b7 2017-07-21 [egs] update hkust results (#1772)

5.2.73 2c51a3ff1 2017-07-21 [egs] Graphemic lexicon updates / fixes in babel/s5d recipe and hub4_spanish recipe (#1740)

5.2.74 5c3c142b6 2017-07-23 [egs] Add missing make_corpus_subset.sh in babel_multilang example (#1766)

5.2.75 dbdd284f6 2017-07-24 [scripts] Fix convert_nnet2_to_nnet3.py (#1774)

5.2.76 1fef98995 2017-07-26 [src] make dithering in feature processing more efficient by using random state, thanks: liximin244@gmail.com

5.2.77 d79e1c0e4 2017-07-26 [scripts] update nnet3 scripts to fix bug where rejecting 'bad' models was not happening (#1777)

5.2.78 caaf6fab2 2017-07-31 [tools] adding phonetisaurus install scripts (PR#1734)

5.2.79 af4b687a5 2017-07-31 [src] Circumvent Visual Studio 2017 bug regarding name resolution (#1783)

5.2.80 ff2ad5012 2017-08-02 [egs] update multi_condition script in swbd (#1788)

5.2.81 35faa4d83 2017-08-03 [src,egs,scripts] Cosmetic and other minor fixes, some required for segmentation PR (#1784)

5.2.82 c4a73526b 2017-08-04 [egs] fix bugs in Multi-database English LVCSR recipe (#1785)

5.2.83 3f61293f2 2017-08-04 [egs] Update tedlium and aspire recipes to use xconfigs (avoid crash) (#1790)

5.2.84 e0a82e424 2017-08-05 [egs] Fix bug RE xent_regularize in Aspire chain recipes. (#1797)

5.2.85 5e91c5d2e 2017-08-07 [src] Cosmetic fix to usage message (#1800)

5.2.86 53e5e1221 2017-08-08 [egs] Add chain recipe for Fisher English (#1803)

5.2.87 4d27debb7 2017-08-09 [src] Fix bug in model-update consolidation code (thanks: sriram ganapathy).

5.2.88 4a0106a75 2017-08-09 [src] Small test-code fix

5.2.89 cc2469e86 2017-08-09 [src] modify feature-extraction binaries to support downsampling (#1773)

5.2.90 c216b5403 2017-08-10 [scripts] reverting the recent change of perturb_data_dir_speed_3way.sh (#1809)

5.2.91 1a508c3a1 2017-08-11 [egs] A fix to swbd1_data_prep.sh to make speaker ids prefixes of the utterances; a change in WSJ script to compute the PCA instead of LDA. (#1810)

5.2.92 8e269d04c 2017-08-11 [build,scripts] Fix to fast-lstm xconfig layer which was not functional; cosmetic change to 'configure'.

5.2.93 aaf680e55 2017-08-11 [src,scripts,egs,build] Enable RNNLM lattice rescoring with Tensorflow (#1741)

5.2.94 d117a6422 2017-08-12 [scripts] Documentation fix in xconfig scripts

5.2.95 511350fff 2017-08-14 [scripts] Fix to script usage message (thanks: @yzmyyff)

5.2.96 66bcdceed 2017-08-14 [build] fix compilation problem of tfrnnlm and tfrnnlmbin (#1822)

5.2.97 12fa929fc 2017-08-14 [scripts,src] Check that symbol '#0' is not in the vocab of the ARPA LM file or the lexicon [thanks: nitzan@almagu.com] (#1806)

5.2.98 5846a2d13 2017-08-15 [src] Inconsequential bug-fixes to problems found when compiling with -DKALDI_PARANOID

5.2.99 eaebe0893 2017-08-15 [build] Enable Travis cache for faster Travis builds, avoiding timeouts (#1819)

5.2.100 6c6d1dd42 2017-08-17 [build] Change to compilation flags for Sequitur on Mac (#1716)

5.2.101 87b55e10c 2017-08-17 [scripts,egs] fixed issues in multilingual training with --use-ivector option (#1828)

5.2.102 3b185c715 2017-08-17 [build] remove debug print (fixes some of the travis builds failures) (#1827)

5.2.103 4816ee5f2 2017-08-18 [scripts] fix lmrescore_rnnlm_lat.sh: fstproject doesn't work on const-arpa FST (#1830)

5.2.104 91d47c1dc 2017-08-20 [doc] Fix typos (#1833)

5.2.105 92dad5b65 2017-08-21 [doc] Add some terms to glossary

5.2.106 0e5bb81e2 2017-08-21 [scripts,egs] Nnet3 segmentation using labels from ASR (#1676)

5.2.107 6d885ee2c 2017-08-23 [scripts] Adding Swbd speech activity detection scripts missing from #1676 (#1838)

5.2.108 e89280576 2017-08-23 [egs] add swbd/s5c/conf/mfcc_hires_bp.conf missing from local/run_asr_segmentation.sh (#1841)

5.2.109 e8636e910 2017-08-24 [src] fix #include directive in compressed-matrix.h, include subfolder. #1842 (#1843)

5.2.110 d234b076a 2017-08-24 [egs] Note small bug in AMI run.sh, thanks @akreal

5.2.111 1d1373130 2017-08-24 [scripts] fixes to nnet3-based segmentation scripts (#1844)

5.2.112 e5a48fc0e 2017-08-25 [src] fixed copy-constructor bug RE test_mode option for DropoutComponent (#1845)

5.2.113 ce6b342d5 2017-08-25 [scripts,src,egs] simplify by removing --leftmost-questions-truncate option (#1840)

5.2.114 09554c614 2017-08-28 [egs] Aspire example scripts: Update autoencoder example to xconfig (#1847)

5.2.115 4af4f3699 2017-08-28 [src] Remove repeated parameter documentation. (#1849)

5.2.116 7246f2cb9 2017-08-29 [scripts] bug-fix in TFRNNLM rescoring script (no 'ark' needed for unk.probs file) (#1851)

5.2.117 bca082ed9 2017-08-29 [egs] Update to score.sh in fisher_swbd setup, allow --iter option (#1853)

5.2.118 d8e42b07d 2017-08-29 [src] Typo fixes in documenation. (#1857)

5.2.119 355f4e169 2017-08-30 [src] Adding documentation for lattice discriminative training functions (#1854)

5.2.120 807dc486c 2017-08-30 [scripts] Add --decode-extra-opts in steps/decode.sh (required for speech activity detection scripts) (#1859)

5.2.121 018e79edb 2017-08-31 [doc] Fix to how documentation is built (thanks: David van Leeuwen)

5.2.122 683e3b565 2017-09-01 [scripts] Fix default for egs.cmd in nnet3 training scripts (#1865)

5.2.123 8c792f0cb 2017-09-01 Remove memory leaks and unused variables (when CUDA is not enabled) (#1866)

5.2.124 0a8fdbfa5 2017-09-05 [build] Cosmetic fix in Makefile

5.2.125 707483088 2017-09-05 [egs] Fixed some issues in the multilingual BABEL example scripts (#1850)

5.2.126 f4a6a664f 2017-09-08 [scripts] Bug-fix in long-utterance-segmentation script (thanks: Armin Oliya) (#1877)

5.2.127 134cf7ee5 2017-09-08 [doc] Keep tutorial code up to date (thanks: Luwei Yang)

5.2.128 26e8b5616 2017-09-11 [scripts] A cosmetic change to info messages in chain training (#1880)

5.2.129 8ad898c2b 2017-09-12 [build] dependency-check: that python2.7 and python3 exist and 2.7 is default (#1876)

5.2.130 ca871fb25 2017-09-13 [egs] Fix inconseqential input-checking bug in Swbd example script (#1886)

5.2.131 1771a2543 2017-09-15 [src] Fix bug in block matrix addition (thanks: Sidhi Adkoli).

5.2.132 d1016d8df 2017-09-15 [src,scripts,egs] Attention modeling, with example scripts (#1731)

5.2.133 82686ee6b 2017-09-15 [src,scripts,egs] Transfer learning for ASR with nnet3 (#1633)

5.2.134 cddb72ce0 2017-09-17 [doc] Fix typo in the About page (#1890)

5.2.135 ecd48ca7f 2017-09-17 [build] Updates related to refactoring of the phonetisaurus build system. (#1891)

5.2.136 633310c0d 2017-09-22 [src] Fix CU_SAFE_CALL wrapper so it correctly prints CuBLAS error codes (#1900)

5.2.137 13e3a0be2 2017-09-22 [src,scripts] Remove some unused code and scripts (#1904)

5.2.138 9a6e1c070 2017-09-22 [egs,scripts] For scoring analysis scripts, guess if it's utf-8 compatible stream or not (#1903)

5.2.139 d4bb00a9b 2017-09-25 [egs] Speaker recognition scripts for FAME! Speech Corpus (#1897)

5.2.140 69e3c49de 2017-09-26 [build] Remove invalid dependency in Makefile (thanks: @qzfnihao)

5.2.141 b5d7d41e8 2017-09-26 [scripts] small fix for utils/dict_dir_add_pronprobs.sh (#1912)

5.2.142 b0cc15790 2017-09-26 [scripts] Change hostname -d to hostname -f for mac compatibility.

5.2.143 906006f5a 2017-09-27 [scripts] Fixes to data-cleanup scripts (#1902)

5.2.144 ba00b18c2 2017-09-27 [build] make configure script allow cuda 9.0 (#1916)

5.2.145 6cab750e8 2017-09-28 Modify data-validation script and dictionary-validation script to disallow exotic space characters (#1910)

5.2.146 fad21abdf 2017-09-28 [build] remove obsolete arch from cuda gencode (#1919)

5.2.147 87107ed4c 2017-10-03 [egs] fix run_ivector_common.sh in some setups to use data/lang not data/lang_nosp (#1921)

5.2.148 f348b26c1 2017-10-03 [build] Fix TensorFlow installation issues (#1920)

5.2.149 e082c17d4 2017-10-03 [src,scripts,egs] Xvectors: DNN Embeddings for Speaker Recognition (#1896)

5.2.150 0f2df8014 2017-10-04 [egs] Updating WSJ TDNN example to use batchnorm instead of renorm.

5.2.151 9eccdbebc 2017-10-09 [src] Some fixes to cu-rand to ensure deterministic behavior. (#1927)

5.2.152 98757696e 2017-10-09 [scripts] Minor fix to nnet3 training scripts RE log parsing/time-reporting (#1926)

5.2.153 5fbdc666d 2017-10-10 [scripts] Fixing OS X compatibility issue for tabs with sed. (#1928)

5.2.154 971fe628c 2017-10-11 [egs] Replace Zenkaku-Space to Zenkaku-Underscore in CSJ recipe (#1931)

5.2.155 8bb27a921 2017-10-11 [src] Loosen a threshold in the derivative test to make it pass (#1936)

5.2.156 7ed731112 2017-10-11 [scripts] Replace sed commands using \t and \n for OS X compatiblity (#1935)

5.2.157 e3f2e0ca7 2017-10-11 [egs] minor script fix; remove unused variable (#1933)

5.2.158 fe6f86b06 2017-10-11 [egs] In csj setup, use KALDI_ROOT instead of ../../.. (#1932)

5.2.159 e3fb7c4a0 2017-10-11 [egs] Small fixes to egs/sre16/ and egs/bn_music_speech dataprep, link to pretrained models (#1934)

5.2.160 696696b6f 2017-10-13 [scripts] Fix small script typo (cleanup-related); thanks, @francoishernandez.

5.2.161 510428f86 2017-10-13 [egs] Adding new non-splicing layers in some TDNN setups. (#1937)

5.2.162 9e37c5178 2017-10-16 [src,egs] Minor tweaks in nnet3 code (diagnostics,testing); add stage option to swbd setup.

5.2.163 63774a542 2017-10-17 [scripts,egs] Use KALDI_ROOT instead of relative path (#1941)

5.2.164 167c2b904 2017-10-17 [doc] Fix path in kaldi for dummies (thanks: Ermand Hoxha)

5.2.165 fa6bee362 2017-10-19 [egs] Librispeech: make sure phone list is sorted, for repeatability (#1946)

5.2.166 20f81b5dc 2017-10-19 [egs] Adding new non-splicing layers in the AMI TDNN setup (#1947)

5.2.167 c96fb5d5e 2017-10-19 [scripts] Add new tuning option for speech activity detection (SAD) scripts (#1949)

5.2.168 e4952cfe4 2017-10-19 [egs] Fixed a missing xconfig change of the tdnn recipe in the previous commit (#1948)

5.2.169 020c50089 2017-10-19 [scripts] Simplify and remove non-useful diagnostics in nnet3 training scripts

5.2.170 edb1aae94 2017-10-20 [egs,scripts] Disambiguate cmd.sh and path.sh pathnames: add './' (#1953)

5.2.171 3c7f0ddb2 2017-10-21 [egs] fix to scoring script error in HKUST recipe (#1957)

5.2.172 f89ffeb6b 2017-10-22 [src,scripts,egs] Make it possible to specify component-level l2 regularization (#1959)

5.2.173 4006d9466 2017-10-23 [egs] Updates to Chime4 recipe (#1956)

5.2.174 582993a43 2017-10-24 [scripts] Small fix to nnet3 training scripts: pass egs_command to get_egs.sh (#1970)

5.2.175 de247358d 2017-10-24 [egs,scripts] Add support for l2-regularize to LSTM layers; give example.

5.2.176 e8e86ae0d 2017-10-24 [egs] Small egs fixes to HKUST and sre10 (#1958)

5.2.177 d38b5c5bb 2017-10-25 [egs] Further small HKUST and SRE fixes (#1971)

5.2.178 fb3778740 2017-10-25 [build] Make nvcc respect CXX (#1965)

5.2.179 34c3e2ca5 2017-10-25 [src,scripts] nnet1: various minor changes, (#1962)

5.2.180 41301bdda 2017-10-25 [scripts] Fix a bug in attention xconfig regarding key-scale (thanks: Pegah) (#1975)

5.2.181 ddf71be1f 2017-10-27 [build] Remove download of ATLAS header files from tools/. (#1974)

5.2.182 81bd30a28 2017-10-27 [egs] Small fix to Chime4 RE data location (#1966)

5.2.183 9d7247bd3 2017-10-29 [egs] Add example of component-level l2-regularize for WSJ scripts

5.2.184 323100178 2017-10-29 [src] Minor fix: change to error message (#1980)

5.2.185 13333d1cc 2017-11-01 [scripts] Fixes to segment_long_utterances.sh (thanks @christophfeinauer) and train_raw_dnn.py (#1993)

5.2.186 0b818c442 2017-11-02 [egs] Remove deprecated non-working scripts

5.2.187 3ea534070 2017-11-02 [egs] Minor fixes to the SRE16 v2 recipe (#1986)

5.2.188 148c884b7 2017-11-02 [egs] rotation for image augmentation in CIFAR example (#1955)

5.2.189 c0c213f4d 2017-11-03 [src,scripts] nnet1-related changes: (#1998)

5.2.190 38296ad2f 2017-11-04 [egs] fisher_callhome_spanish: updated phonetic transcription and scoring; other fixes (#1992)

5.2.191 978e38ee9 2017-11-06 [scripts] Fix edge cases (#1994,#1991) in data-cleanup scripts (#1997)

5.2.192 e003f2925 2017-11-07 [egs] Small fix to tedlium score_basic.sh (which is not used by default), thanks @danoneata

5.2.193 b01b5b41a 2017-11-07 [src] Fix to nnet training to print multiple objfs in deterministic order (and print all of them) (#2001)

5.2.194 d3dd54553 2017-11-07 [egs] Updating multi_en recipe's data selection, training procedure and lexicon preparation,.etc (#1893)

5.2.195 221664ea4 2017-11-08 [build] Tighten .gitignore exclude rules to ignore dot files in /src/*/ (#2004)

5.2.196 26be81eed 2017-11-09 [egs,src] Some inconsequential fixes (#2007)

5.2.197 8569f8df3 2017-11-09 [egs] BABEL scripts: new config files, and config-file generator script (#2006)

5.2.198 061d12682 2017-11-09 [scripts,egs] Convert tab->space, and indent correctly (#2008)

5.2.199 05b2aed18 2017-11-10 [egs] cleanup of BABEL scripts (#2011)

5.2.200 b952cf33a 2017-11-10 [scripts] Documentation fixes. Thanks: Rongjin Li.

5.2.201 640a8e3a1 2017-11-13 [egs] Update soft link run_tdnn.sh in WSJ example

5.2.202 ac8adf2cd 2017-11-13 [egs] minor fixes to multi_en recipe (#2013)

5.2.203 030833abb 2017-11-14 [egs] fix xent parameter (formal change, no effect on results) (#2014)

5.2.204 1d66675e0 2017-11-14 [egs] Babel cleanup + updated to ivector training with pca-transform (#2016)

5.2.205 0884818d2 2017-11-15 [src] nnet1: adding 'multitask' loss to 'nnet-train-perutt', (#2017)

5.2.206 f19e94c21 2017-11-18 [scripts,doc] Add check for tilde in wav.scp (thanks: tessfu geteye); update queue documentation.

5.2.207 9e7644cac 2017-11-19 [src] Minor fix: useless decoder declaration removed (#2020)

5.2.208 9b86ff225 2017-11-20 [egs] new BABEL DNN training recipes (#2021)

5.2.209 d508904e8 2017-11-20 [build] add mmseg to .gitignore (#2024)

5.2.210 54749fd07 2017-11-20 [scripts] fail when the directory does not have the right files (#2023)

5.2.211 daeac30bb 2017-11-21 [src] Fix KaldiMatrix::Resize silent failure when resize_type == kCopyData (#2025)

5.2.212 23586f5f7 2017-11-21 [build,src] Increase OpenFst version from 1.6.2 to 1.6.5 (#2026)

5.2.213 d66906dbb 2017-11-21 [scripts] minor fix: make ivector id checking more robust (#2028)

5.2.214 9c877df42 2017-11-21 [egs] babel: fix the lithuanian pitch accent handling (#2029)

5.2.215 2f90f8155 2017-11-22 [build] Update Windows installation (#2032)

5.2.216 5e7deb91a 2017-11-22 [egs] babel: cosmetic change (print only one path when generating language configs) (#2034)

Version 5.3

Major changes that were made between the end of 5.2.x and the start of the 5.3 branch include:

  • Create a nnet3-based setup for RNN language models (i.e. recurrent and neural net based language models)
  • Some extentions to the core of the nnet3 framework to support constant values and scalar multiplication without dedicated compoennts.

Below are commits corresponding to minor version numbers 5.3.x.

5.3.0 db2865034 2017-11-23 Kaldi 5.3 (includes nnet3-based RNNLM setup) (#1693)

5.3.1 39c8d141b 2017-11-24 [scripts,egs] replace all instance of "gensub" with gsub or sub for compatibility (#2036)

5.3.2 0ba0b124c 2017-11-26 [src,egs] fix to fgmm-global-acc-stats-post so that posteriors are read sequentially and features are random access (rather than vice-versa) (#2039)

5.3.3 e96eec1f6 2017-11-26 [src] nnet2: nnet-am-compute, adding option --divide-by-priors (default=false) (#2041)

5.3.4 1f5d747fe 2017-11-27 [build,src] Increase configure version (and add functions to CuVector) (#2042)

5.3.5 bafb68c17 2017-11-27 [src] Fix to run_lstm.sh in swbd/s5c setup: mdeish@gmail.com

5.3.6 9effa4a1e 2017-11-27 [src] add max_warnining option to SlidingWindowCmnOptions to limit the number of times the variance flooring warning is printed (#2043)

5.3.7 92c99ee51 2017-11-27 [egs] Modify some storage paths for striped data (#2038)

5.3.8 6d6e7a908 2017-11-27 [src] Various nnet3 changes, see PR for details (#2031)

5.3.9 acca2ac6d 2017-11-27 [src] Fixes/improvements in error path of TokenHolder code (#2044)

5.3.10 783ad7b26 2017-11-29 [egs] babel: remove some files no longer used (#2048)

5.3.11 35288fae9 2017-11-29 [src] Mostly cosmetic changes in error handling (#2047)

5.3.12 fbef11bb2 2017-11-29 [egs] replaced proportional-shrink with l2-regularize in the latest ami tdnn and tdnn-lstm recipes (#2045)

5.3.13 7a8f47100 2017-11-29 [src,egs,scripts] removed extract-rows, create-split-from-vad and some long-unused scripts (#2049)

5.3.14 e9abbff99 2017-11-29 [src] Fixing comments for BatchNormComponent (#2050)

5.3.15 f14e08369 2017-11-30 [build] fix TF compilation issues; fix version of TF to avoid future issues (#2052)

5.3.16 fd5e33608 2017-12-01 [src] Fix "warning: enumeration value ‘CUSPARSE_STATUS_ZERO_PIVOT’ not handled in switch [-Wswitch]" (#2054)

5.3.17 77ae8fe02 2017-12-02 [src] Bug-fix to conceptual bug in Minimum Bayes Risk/sausage code. Thanks:@jtrmal (#2056)

5.3.18 85fa84c91 2017-12-02 [scripts] Adding --num-threads option to ivector extraction scripts; script fixes (#2055)

5.3.19 a3441c243 2017-12-03 [src] Fix some small typos (#2060)

5.3.20 72dfed026 2017-12-04 [scripts] Fix bug in slurm.pl affecting log format (#2063)

5.3.21 93ceca700 2017-12-04 [egs] Add recipe for Mozilla Common Voice corpus v1 (#2057)

5.3.22 8bc989982 2017-12-04 [doc] Add documentation about matrix row and column ranges in scp files.

5.3.23 cd6562753 2017-12-04 [scripts,egs] Convert ". path.sh" to ". ./path.sh" (#2061)

5.3.24 c9481387a 2017-12-04 [egs] Update to CSJ example scripts, with chain+TDNN recipes. Thanks: @rickychanhoyin (#2035)

5.3.25 8cc5c8b32 2017-12-05 [src] Fix compilation error on MSVC: missing include. (#2064)

5.3.26 87783fc4e 2017-12-08 [src] Small fix RE CuSparse error code printing (#2070)

5.3.27 41806f648 2017-12-09 [src,egs] Small cosmetic fixes (#2074)

5.3.28 cc7c2f6d7 2017-12-10 [egs] Add a tdnn+lstm+attention+backstitch recipe for tedlium (#1982)

5.3.29 f51f8ca25 2017-12-11 [src] Bug-fix to assertion in cu-sparse-matrix.cc (RE large matrices) (#2077)

5.3.30 f5409927e 2017-12-11 [egs] Add an l2-regularize-based recipe for image recognition setups (#2066)

5.3.31 4e3c1832a 2017-12-12 [egs,scripts] Add OPGRU scripts and recipes (#1950)

5.3.32 5799dae14 2017-12-13 [scripts] Fix augment_data_dir.py (relates to non-pipe case of wav.scp) (#2081)

5.3.33 eaa6a26a2 2017-12-13 [egs] Small fix to backstitch in AMI scripts (#2083)

5.3.34 23e046398 2017-12-14 [egs] Copy dictionary-preparation-script fix from fisher-english(8e7793f) to fisher-swbd and ami (#2084)

5.3.35 e587c15af 2017-12-15 [egs] Mac compatibility fix to SGMM+MMI: remove -T option to cp (#2087)

5.3.36 2ec515d78 2017-12-16 [src] Fix bug in nnet3 optimization, affecting Scale() operation; cosmetic fixes. (#2088)

5.3.37 e7fe053f8 2017-12-19 [scripts,egs] Some fixes to OPGRU scripts from Lu Huang (#2091)

5.3.38 04458f7a2 2017-12-20 [src,egs,scripts] Merging RNNLM-related changes which were in wrong branch (#2092)

5.3.39 a3b3bfaf6 2017-12-21 [src] Prevent RNNLM training from exhausting GPU memory in diagnostic jobs (#2094)

5.3.40 f61224fa7 2017-12-21 [src] Fix bug in compilation with Scale() expressions. (#2098)

5.3.41 81b7a1947 2017-12-21 [egs] Fix to decode context option in swbd (#2100)

5.3.42 3c21ef23f 2017-12-22 [doc] Fix typo in docs (#2101)

5.3.43 d46ff7e82 2017-12-26 [src] Ensure num-tree-leaves is a multiple of 8 (RE NVidia Volta architecture) (#2097)

5.3.44 48656c314 2017-12-26 Revert "[src] Ensure num-tree-leaves is a multiple of 8 (RE NVidia Volta architecture) (#2097)" (#2105)

5.3.45 a5561c3cf 2017-12-26 [src,scripts] Simplify model combination: do simple average over last n models (#2067)

5.3.46 2b3acc585 2017-12-27 [scripts] Cosmetic change to validate_data_dir.sh

5.3.47 843bc207b 2017-12-28 [egs] Add TDNN recipes (and other changes) for chime4 (#2012)

5.3.48 0b4ea277f 2017-12-28 [src] Dropout and batchnorm fixes for discriminative training (#2109)

5.3.49 df7a41978 2017-12-28 [scripts] Fix validation-script bugs found by Somogyi Z; make validate_text.pl quieter.

5.3.50 47ea76bd1 2017-12-28 [egs] Update swbd/s5c RNNLM setup with hesitation mapping, and perplexities (#2108)

5.3.51 84f9b86a2 2017-12-28 [egs] fisher_swbd chain script bug fix (thanks: Filip Jurcicek) (#2110)

5.3.52 12706348f 2017-12-29 [egs,src] Fix cleanup script bug (thanks: CW Huang), clarify comment.

5.3.53 9ac5dd4ef 2017-12-31 [src] Make faster ApplyFloor and ApplyCeiling variants for Matrix and CuMatrix. (#2115)

5.3.54 006e3c8cc 2017-12-31 [scripts] Fig bug in validate_data_dir.sh introduced in df7a41978f2. Thx:@jcsilva

5.3.55 1431db17e 2017-12-31 [scripts] Fix lattice_oracle_align.sh bug (Thx: @roman-verbit-ai)

5.3.56 c08a1c197 2018-01-01 [build] Add new search dir for ATLAS (Thx: Sari Sultan)

5.3.57 6f9fab72f 2018-01-02 [scripts] Fix script issue affecting some xvector training (thanks: daniel garcia-romero)

5.3.58 b808365a6 2018-01-03 [egs] Improve comments for fisher_swbd tdnn_lstm_1a (#2122)

5.3.59 8292e4cbd 2018-01-04 [egs] Add OCR/Handwriting Recognition examples (#1984)

5.3.60 33db55aeb 2018-01-04 [egs] swbd/s5c: Add RNNLM setup for where Fisher text data not there. (#2118)

5.3.61 36df91d35 2018-01-05 [src] Round number of leaves to multiple of 8, will be faster with Volta architectures (#2106)

5.3.62 40a7d70d0 2018-01-06 [src] test-code change for speedup (#2128)

5.3.63 42767abec 2018-01-06 [scripts] Script fix in RNNLM setup; cosmetic change.

5.3.64 fd9284ac4 2018-01-07 [scripts] Add support for grayscale images to matrix_to_image.py (#2131)

5.3.65 e7ddb06b2 2018-01-07 [egs,scripts] source cmd.sh in RNNLM setups; rename some RNNLM setups; script fix (#2132)

5.3.66 33b0de92c 2018-01-07 [src,scripts,egs] Add n-best rescoring for Kaldi-RNNLM (#2120)

5.3.67 a0b71317d 2018-01-09 [src] Change matrix/ include structure to avoid leaking global symbols like 'real' (#2127)

5.3.68 360c96ca4 2018-01-10 [src] Temporary patch for build-tree assertaon failure.

5.3.69 e7e488edc 2018-01-10 [egs] updating fisher_swbd nnet3/chain scripts: removing non-chain examples (#2136)

5.3.70 da18a45b6 2018-01-10 [src] Fix bug in rounding num-leaves (#2145)

5.3.71 116a948a3 2018-01-10 [build] Avoid problems when installing pocolm/irstlm a 2nd time (#2144)

5.3.72 e1ead787a 2018-01-10 [scripts] Cosmetic typo fix (#2143)

5.3.73 65b856f73 2018-01-11 [egs] Add Spanish recipe based on Heroico corpus (#1882)

5.3.74 6fb73b024 2018-01-12 [scripts] Extend num-digits in get_uniform_subsegments.py.. avoid crash on long utts (#2148)

5.3.75 5a8b4a605 2018-01-12 [build] update configure and android_openblas.mk for the latest ndk in android. (#2124)

5.3.76 26df3d818 2018-01-12 [src] Bug-fix in nnet3 compilation, RE Scale() expressions

5.3.77 1fb7fd92e 2018-01-15 [scripts] OS X compatibility fix to utils/format_lm.sh, thx: @johnhosephmorgan

5.3.78 d883e305f 2018-01-15 [build] set 'rpath' for ATLAS shared libraries, (#2153)

5.3.79 9e6dbab70 2018-01-17 [egs] ami, changing path for storing downloaded wavs, (#2154)

5.3.80 26b0746f0 2018-01-17 [src,scripts,egs] ivector-based diarization recipe for CallHome (#1894)

5.3.81 829b013db 2018-01-17 [egs] Update to CSJ example: fixed typo and updated to use xconfig (#2160)

5.3.82 687d31366 2018-01-18 [src] Fix bug when compiling with flag -DNDEBUG (#2162)

5.3.83 c739ff95b 2018-01-19 [egs] Minor fix in Farsdat recipe RE sctk path (#2164)

5.3.84 85a2c60e3 2018-01-20 [src] Print informative error if num-ceps >= num-mel-bins in MFCC (#2166)

5.3.85 35950ea24 2018-01-21 [egs] add speaker recognition recipe for aishell (#2111)

5.3.86 d6391f864 2018-01-22 [scripts,egs] Fix bug in slurm.pl (thx:@kamo-naoyuki), remove outdated results.

5.3.87 96cbdd697 2018-01-22 [egs] Update chain results and add new scripts for fisher_swbd (related to #2136) (#2173)

5.3.88 8596bbf54 2018-01-23 [egs] removing broken link in babel multilang setup (#2177)

5.3.89 476cb3f6b 2018-01-23 [egs] fixes for the IAM example (#2176)

5.3.90 b73bb12e4 2018-01-23 [egs] Make sure scoring opts are passed to score_cer.sh in UW3 (#2181)

5.3.91 47de14528 2018-01-23 [scripts] Fix typos in scripts (#2182)

5.3.92 2e105fc45 2018-01-23 [scripts] Fix to get_num_frames.sh for large datasets, RE truncation in awk (#2174)

5.3.93 7cf434ce1 2018-01-23 [scripts] Fix python3 compatibility bug (#2184)

5.3.94 ed84a5107 2018-01-24 [scripts] Prevent crash when input_model is None (#2188)

5.3.95 6fed4c7dd 2018-01-24 [src] Remove CuDevice destructor (avoid cuda-memcheck warnings) (#2185)

5.3.96 7ee7893f9 2018-01-24 [src] Fix nnet3 back-compatibility issue (thx: @satmass)

5.3.97 f861b00ca 2018-01-25 [src] Speed fix to online decoding (thanks: David van Leeuwen) (#2180)

5.3.98 3031e26e6 2018-01-25 [src] Optimization to decoders for speed (#2168)

5.3.99 0f14373f4 2018-01-28 [src] Work around problem related to ungetc failures on ifstream (#2194)

5.3.100 30e9a90d3 2018-01-28 [src] Small fix to component reading, workaround for ungetc() issue.

5.3.101 e4fc87d14 2018-01-30 [scripts] bugfix for 'steps/cleanup/clean_and_segment_data.sh', (#2196)

5.3.102 1647856ee 2018-01-31 [egs] Small bug-fix in Librispeech recipe (#2190)

5.3.103 2de3b3866 2018-01-31 [src] Make arpa2fst robust against ARPA files without <s>. (#2167)

5.3.104 c82560ddb 2018-02-01 [scripts] Fixed small issue get_uniform_subsegments.py (RE rounding) (#2200)

5.3.105 b4fbe00b0 2018-02-02 [egs] Add assert to check --backstitch-training-interval option (#2203) (#2204)

5.3.106 9e2d8442b 2018-02-03 [egs,scripts] Fix and simplify speed-perturbation scripts; fix permissions. (#2205)

5.3.107 79065901b 2018-02-06 [doc] remove outdated TODOs. Thx: David Van Leeuwen.

5.3.108 027f1d796 2018-02-07 [scripts] Add missing import statement in script (#2207)

5.3.109 8e170e039 2018-02-07 [egs] multi_en: Fixed acronym normalization, swbd lexicon preparation, OOV … (#2137)

5.3.110 f2ab7d403 2018-02-08 [src] Fix small bug in Log1p (thanks: Max Lvov)

5.3.111 1f1b1089c 2018-02-08 [scripts] make format_lm.sh work when source == target directory (#2209)

5.3.112 c52ee4c52 2018-02-09 [src] rnnlm-embedding-training bug fix (thanks: @yanglin187) (#2208) (#2210)

5.3.113 ca73b1f5a 2018-02-10 [scripts] Fix warning from create_data_links.pl (#2212)

5.3.114 611aa612c 2018-02-11 [scripts,egs] Fix failures in some run_ivector_common.sh script introduced in #2205 (#2213)

5.3.115 8a384379b 2018-02-11 [egs,scripts] Recipe for voice activity detection on Switchboard (#2193)

5.3.116 ca5f128f2 2018-02-11 [doc] Updating contributor list in COPYING to contain the current top 66 contributors in Kaldi. Thanks to Yishay Carmiel for gathering this info (#2214)

5.3.117 5ea9b0da3 2018-02-13 [egs,scripts] HUB4 English Broadcast News recipe (#2027)

5.3.118 6d32a0bbf 2018-02-14 [src] Fix usage message, comment of ali-to-pdf (#2218)

Version 5.4

The main changes that were made between the end of 5.3.x and the start of the 5.4 branch include:

  • Some code changes in the nnet3 codebase, for speed and memory efficiency.
  • Various simplifications and code reorganizations in the nnet3 code.
  • Support for a new kind of factorized TDNN (TDNN-F) which gives substantially better results than our old TDNN recipe, and is even better than our old TDNN+LSTM recipe. A good example of this is in egs/swbd/s5c/local/chain/tuning/run_tdnn_lstm_1n.sh. Some nnet3 code changes were needed for this as well (mostly: support for constraining a matrix to have orthonormal rows).

Some of the larger changes that were made while 5.4 was the major version number include:

  • Improvements to handwriting recognition and OCR recipes, including BPE (word-piece) encoding.
  • An updated version of the TDNN-F configuration, including ResNet-style bypass, which is now the default in many recipes. (it's called tdnnf-layer in xconfigs).
  • A rewrite of the CUDA memory allocator to be based on a small number of large regions (since with newer drivers and hardware, allocation speed was becoming a bottleneck).
  • A decoder speedup (make use of OpenFst's NumInputEpsilons() function).

Below are commits corresponding to minor version numbers 5.4.x.

5.4.0 be969d7ba 2018-02-17 [build] Upgrading version of Kaldi to 5.4

5.4.1 6016c004b 2018-02-17 [doc] Add documentation for 5.4 version of Kaldi

5.4.2 9b282b58c 2018-02-18 [scripts] Fix bug in steps/diagnostic/analyze_phone_length_stats.py [thanks: Vinay Sunder.]

5.4.3 09b017669 2018-02-18 [build] Update gst-plugin Makefile (#2226)

5.4.4 b61d76049 2018-02-20 [src] Fix error in how Compress() command is printed for debugging

5.4.5 6dd9c3e3f 2018-02-21 [scripts] Fix bug in steps/libs/nnet3/train/frame_level_objf/common.py (#2231)

5.4.6 e3df6ac0c 2018-02-21 [src] Print curand return codes (#2232)

5.4.7 9191f296d 2018-02-23 [egs] Fix script bug (in error-checking statement) (#2239)

5.4.8 e0d8f13e8 2018-02-23 [scripts] Fix typo (#2237)

5.4.9 4f27915e0 2018-02-23 [scripts] Fix typo in comment (#2238)

5.4.10 f629bf89a 2018-02-24 [scripts] Bug-fix in analyze_phone_length_stats.py (thanks: @stanleyguan)

5.4.11 851c5e4c6 2018-02-27 [egs] Fix to swbd chain tdnn_blstm script bug (thanks: sameer khurana)

5.4.12 7e902f535 2018-02-27 [scripts] Fix for nnet3 model combination on CPU when --use-gpu false (#2240)

5.4.13 27f6c5ed1 2018-02-28 [src] Fix small bug in cu-math.cc affecting non-GPU LSTM training, thx: Mael Primet.

5.4.14 0ebe5b59e 2018-02-28 [egs] Simplify Aspire scripts and make it work with LDC release. (#2230)

5.4.15 9a4ba5ec0 2018-02-28 [src,scripts,egs] Add form of dropout that shares the mask across frames. (#2244)

5.4.16 8b500766f 2018-02-28 [scripts] add --max-arcs option to scripts/rnnlm/lmrescore_pruned.sh (#2245)

5.4.17 c42682602 2018-03-01 [src] Fix to cu-rand.cc RE large matrix sizes. (#2249)

5.4.18 bc0fcad41 2018-03-02 [build]: Dockerfile for kaldi installation (#2246)

5.4.19 e7b3a66cb 2018-03-03 [src] Fix macro redefinition problems in tensorflow rnnlm. (#2250)

5.4.20 03b0ea831 2018-03-03 [src,scripts,egs] Enable backstich training for RNNLM; add examples (#2096)

5.4.21 03edf8248 2018-03-04 [src] Fixed bug in GeneralDropoutComponent when block-dim set (#2252)

5.4.22 0e1822d86 2018-03-04 [egs] Port factorized TDNN script to librispeech (#2233)

5.4.23 230d013b4 2018-03-05 [scripts] python3 compatibility: change / to // (#2253)

5.4.24 4518c7769 2018-03-05 [src] Fix CPU swap methods of CuVector and CuPackedMatrix (#2254)

5.4.25 25391fb66 2018-03-05 [scripts] Add example of accessing nnet3 model from python (#2255)

5.4.26 e74b918dc 2018-03-05 [src,scripts,egs] End-to-end chain training (#2072)

5.4.27 ed05f17fa 2018-03-06 [egs] thchs-30: fix Mac compatibility issue (#2258)

5.4.28 f293ccd23 2018-03-06 [build] upgrade the install_beamformit script (#2259)

5.4.29 65186851e 2018-03-07 [src] Changes to logging information printed by nnet1 (#2261)

5.4.30 9e3ce6749 2018-03-07 [scripts] Fix python3 issues w.r.t. relative imports, in steps/libs/. Thx: Cemil Demir (#2263)

5.4.31 7f028e307 2018-03-07 [egs] Adding orth-factorized-TDNN recipe for fisher_swbd (#2264)

5.4.32 c541ea017 2018-03-08 [src] Fix small bug in nnet3 component (config was lost on read).

5.4.33 b4711ada5 2018-03-08 [src] Fix small bug in nnet3 component (config was lost on read)... 2nd try.

5.4.34 0360215ff 2018-03-09 [src] Support floating orthonormal-constraint (#2267)

5.4.35 19dc26ff8 2018-03-12 [src] bug-fixes for end2end chain code (#2270)

5.4.36 be8f4ec44 2018-03-12 [src] Fix subsampling factor in nnet3 egs (#2265)

5.4.37 5eb57ccb7 2018-03-13 [egs] Add chime5 baseline recipe (#2262)

5.4.38 6363d5520 2018-03-13 [scripts] Add delta_opts for train_mono.sh (#2221)

5.4.39 060d92aba 2018-03-14 [egs] fix generate_plots.py to plot deriv-avg and oderiv-rms (#2274)

5.4.40 f1c927c87 2018-03-14 [tools] fix beamformit installation (#2276)

5.4.41 ec98e713f 2018-03-14 [scripts] Bug-fix in convert_data_dir_to_whole.sh (thx: @olix20) (#2277)

5.4.42 b7f9446d2 2018-03-14 [egs] Add modernized vystadial_cz recipe (#2278)

5.4.43 49a90e062 2018-03-15 [src] Fix to nnet3 looped compilation bug (thanks: Alim Misbulah) (#2279)

5.4.44 8044adacc 2018-03-16 [scripts] Cosmetic fix to prepare_lang.sh (#2284)

5.4.45 8ab6e5395 2018-03-16 Revert "[src] Fix to nnet3 looped compilation bug (thanks: Alim Misbulah)" (#2285)

5.4.46 1644312b8 2018-03-17 [src] Add a nnet3 optimization that tries to replace commands ending in Multi with other commands. (#2229)

5.4.47 322f698a1 2018-03-17 [scripts] Fixes to segmentation scripts (#2273)

5.4.48 df9d16c6c 2018-03-18 [scripts] Add support for --use-gpu=wait to nnet3 scripts (#2289)

5.4.49 5562583a8 2018-03-18 [egs] Add improved TDNN-F example for Switchboard. (#2291)

5.4.50 d2a2909a1 2018-03-19 [src] Enable nnet3-compute to optionally subtract priors (#2294)

5.4.51 9654a7c31 2018-03-19 [build] Remove outdated compute architecture. Thx: @CodesFarmer

5.4.52 c6b3588ed 2018-03-19 [src] Fix to nnet3 compilation issue affecting BLSTMP. Thanks: Alim Misbullah (#2286)

5.4.53 eba50e442 2018-03-20 [src] Enable reading wavs produced on iPhones with extra RIFF tags. (#2293)

5.4.54 22fbdd969 2018-03-21 [egs] fixing LM issue in CHiME5 (#2298)

5.4.55 954e13cb5 2018-03-22 [src] Make ClipGradientComponent zero gradients when clipping threshold is zero (#2301)

5.4.56 dbd513c70 2018-03-22 [src] Make CachingOptimizingCompiler thread safe. Thx: Arseniy Gorin (#2288)

5.4.57 f4a5667df 2018-03-22 [src] Fix to comment

5.4.58 9ae3eb78a 2018-03-22 [src,scripts] Make cache size configurable for xvector extraction (#2290)

5.4.59 6dbe7909d 2018-03-22 [src] Speedup and code simplification for chain supervision merging (etc.) (#2302)

5.4.60 8b415fe52 2018-03-24 [egs] remove redundant step from chime5 recipe (#2306)

5.4.61 9f7e55a0a 2018-03-24 [src] Fix issue with CUDA device initialization if 'wait' specified. Thx: @olix20 (#2295)

5.4.62 e03dd12ec 2018-03-24 [scripts,egs] Fix to perturb_speed_to_allowed_lengths.py; egs fix (thanks: @calderma) (#2307)

5.4.63 cc16eecee 2018-03-26 [windows] fix for compiling on Windows VS2017 (15.5.2) (#2310)

5.4.64 7352760ef 2018-03-27 [egs] minor fix for fisher_swbd scripts and swbd nnet3 scripts (#2316)

5.4.65 e5b6696c0 2018-03-27 [egs] Add end-to-end OCR recipe for IAM (thanks: @aarora8) (#2311)

5.4.66 d7e889028 2018-03-27 [egs] Fix LM/lexicon issues in IAM; Add unk decoding; Update results. (#2315)

5.4.67 785198e3c 2018-03-27 [src] Add some asserts in RNNLM code (#2314)

5.4.68 749839560 2018-03-27 [src] Fix to recent commit RE RNNLM code

5.4.69 8af60bb68 2018-03-28 [src] Apply limits prior to chain denominator computation, avoid failures. (#2308)

5.4.70 5e6bd39e0 2018-03-28 [tools, extras] morfessor installation script (#2299)

5.4.71 191b39a14 2018-03-28 [src,scripts,egs] Semi-supervised training on Fisher English (#2140)

5.4.72 b620633aa 2018-03-30 [build] Updates to Sequitur installation script (for compatibility with new setuptools) (#2323)

5.4.73 e507a7aa3 2018-04-02 [src,egs,doc] Some cosmetic changes

5.4.74 6f4fafbd4 2018-04-03 [src] Bug fix: oscillating cluster threshold in agglomerative-cluster (#2333)

5.4.75 fe36823d8 2018-04-03 [src] Fix bug causing wrong learning rate in RNNLM training with --stage option (#2330)

5.4.76 831dd38cf 2018-04-03 [src] make fst templates inline to eliminate linking errors in other places (#2329)

5.4.77 7436ed690 2018-04-03 [src] Make it so no threads are used if you set num-threads to 0.

5.4.78 a9e4f7430 2018-04-04 [build] Fix some problems with sequitur installation (#2337)

5.4.79 181427856 2018-04-04 [scripts] Make sure nj in get_utt2dur.sh does not exceed num-utts. (thx: @olix20) (#2326)

5.4.80 5294666e8 2018-04-05 [scripts] Fix to nnet3 bug RE per-utt splitting that appeared after #2140; un-support --transform-dir. Thx: @aaror8 (#2334)

5.4.81 1a1e265ae 2018-04-05 [src,build] Fix cygwin compatibility; remove unnecessary memalign macros (#2304)

5.4.82 60862b0e9 2018-04-06 [scripts] Make nnet3 logging output look more like chain logging output (#2338)

5.4.83 c2d796d82 2018-04-06 [scripts] Fix bug in chain/get_egs.sh affecting semi-supervised training (#2339)

5.4.84 c6432957f 2018-04-07 [egs] IAM script fixes using the right lang for training/lattice generation + cleaning (#2340)

5.4.85 c0c55adeb 2018-04-08 [egs] IAM: add chain recipe starting with e2e alignments and tree (#2343)

5.4.86 9b90cbac6 2018-04-09 [egs] Fix scoring issue in IAM (#2346)

5.4.87 e10c4789d 2018-04-10 [src] Revert KALDI_MEMALIGN->aligned_alloc change from #2304, fixes #2347 (#2348)

5.4.88 8a4ab19b5 2018-04-11 [src] Fix nnet3-copy-egs bug for multiling. non-chain training (thx:Joshua Meyer) (#2350)

5.4.89 ad48342ae 2018-04-11 [scripts] Bug fix in combine_egs.sh (for multilingual training) (#2351)

5.4.90 2d0cd7930 2018-04-12 [scripts] Finish removing transform_dir options from nnet3 training (thx: @pstanisl) (#2353)

5.4.91 adcabe40e 2018-04-12 [src] Fix (rare) divergence from orthogonal constraint. Thx: Sergey Tushev. (#2336)

5.4.92 ccd50e293 2018-04-12 [egs] Adding Wellington corpus for LM in IAM OCR (#2354)

5.4.93 e99de506c 2018-04-13 [scripts] Make beam configurable, +default smaller, in pruned RNNLM rescoring (#2325)

5.4.94 535df00c1 2018-04-14 [src] small fix to comment (#2357)

5.4.95 dd7caa6fb 2018-04-14 [egs] A fix to second-stage chain recipes in IAM and lang fixes in UW3 (#2358)

5.4.96 dd1640b0f 2018-04-15 [egs] Fix script bug in IAM setup

5.4.97 4abdb5b48 2018-04-15 [src] Cosmetic code fixes in mel-computations.cc (#2359)

5.4.98 5a48ad47c 2018-04-18 [scripts] Fix to get_saturation.pl, to regex bug introduced around Jan 1 2018. Affects shrinkage of recurrent nnets. (#2371)

5.4.99 8f6bb2d39 2018-04-18 [egs,scripts] Add factorized TDNN training recipe for multi-en; fix python script bug (#2369)

5.4.100 1331a35bc 2018-04-18 [egs] Make Wellington Corpus optional in iam_ocr (#2370)

5.4.101 531810032 2018-04-19 [src,egs] Fix implementation of l2-regularize with backstitch; add ami backstitch example. (#2373)

5.4.102 3040a0ac7 2018-04-20 [egs] Fix commonvoice run.sh (source utils/parse_options.sh to enable --stage). Thx:@jjm (#2375)

5.4.103 9692a61aa 2018-04-20 [src] Fix bug in matrix compression RE stride type. Thx:@LvHang (#2376)

5.4.104 6bd7ca74c 2018-04-20 [src] Decoder fix to bug affecting unusual FST types (#2377)

5.4.105 4fda1b04a 2018-04-22 [src] Comment fix to transition-model.h (#2378)

5.4.106 8ff7fd9f8 2018-04-23 [scripts] Added utils/prepare_extended_lang.sh (extending vocab of a lang-dir) and rnnlm/change_vocab.sh (changing vocab of an existing rnnlm-dir) (#2247)

5.4.107 f0333bb05 2018-04-24 [src,scripts,egs] Support for unconstrained-egs for chain training + example recipes (#2383)

5.4.108 d25b3ef94 2018-04-25 [src] Add a cast to prevent errors from some compilers (#2384)

5.4.109 156ec6c1e 2018-04-25 [egs] Semisupervised training recipes: bug fix regarding context info (#2385)

5.4.110 6a2dce879 2018-04-25 [egs] Remove accidentally committed "exit 0" in WSJ run.sh (#2386)

5.4.111 032c0263c 2018-04-26 [src,scripts] Cosmetic fixes.

5.4.112 5eaf49f57 2018-04-26 [scripts] add wordlists mismatch test for rnnlm rescoring (#2387)

5.4.113 c0ef35615 2018-04-27 [scripts] Clarify usage of diarization make_rttm.py (#2390)

5.4.114 fb54444dc 2018-04-27 [egs] Add speaker verification recipe for the VoxCeleb2 corpus (#2380)

5.4.115 bf40d6103 2018-04-29 [src] Remove un-needed code. Thanks: anand@sayint.ai. (#2393)

5.4.116 5bac98e79 2018-04-29 [egs] Update nnet3/chain scripts for HKUST example (#2368)

5.4.117 e15f3a156 2018-04-30 [scripts] Fix in steps/nnet3/get_egs.sh to avoid failure due to broken pipe (#2397)

5.4.118 1e2a0fd1a 2018-04-30 [egs] Cleanup and update results in IAM OCR scripts (#2395)

5.4.119 64025aa72 2018-05-01 [scripts] rnnlm scripts: ignore first iteration while looking for the best model (#2399)

5.4.120 6a32fbd2e 2018-05-03 [scripts] Fix path.sh bug in vad_to_segments.sh (#2402)

5.4.121 20b2a451e 2018-05-03 [src] Remove unnecessary file chain/context-dep-topology.h (#2401)

5.4.122 080129e97 2018-05-03 [src,scripts,egs] x-vectors for diarization (#2391)

5.4.123 4edec5d0f 2018-05-05 [egs,scripts,src] Add IFN/ENIT dataset Handwriting Recognition Example (#2257)

5.4.124 f9b00e901 2018-05-06 [src] Some small cosmetic changes (#2404)

5.4.125 e18191cba 2018-05-07 [egs] Fix python error by moving __future__ import to top (#2405)

5.4.126 954e69a4f 2018-05-07 [src] make e2e/"unconstrained" numerator computation faster (#2392)

5.4.127 b03b6410f 2018-05-07 [src] Speed up VectorBase<Real>::Sum() by using BLAS (#2394)

5.4.128 08b47be72 2018-05-09 [build] update dockerfile for Ubuntu and cuda (#2408)

5.4.129 90363ea38 2018-05-10 [egs] Update VoxCeleb Recipe (#2403)

5.4.130 bce433657 2018-05-14 [egs] Speeding up i-vector training in voxceleb v1 recipe (#2421)

5.4.131 ff0da26f8 2018-05-14 [scripts] Add util to compute recordings durations before data perturbation (#2414)

5.4.132 f8b678a61 2018-05-15 [scripts] Update to decode_semisup.sh (copy input model) (#2425)

5.4.133 108832dfc 2018-05-17 [egs] Madcat Arabic handwritten text line recognition (#2356)

5.4.134 2ad8d7821 2018-05-18 [egs] Minor fixes to x-vector based recipes (#2426)

5.4.135 fd12973c9 2018-05-18 [egs] update iban tdnn CE and chain recipes (#2429)

5.4.136 558f01023 2018-05-19 [scripts] Add a warning message to nnet3/align.sh about using GPUs. (#2432)

5.4.137 f31c1e93e 2018-05-21 [src,scripts] Support word insertion penalty when getting ctm (#2433)

5.4.138 264a6a3ec 2018-05-21 [scripts] Fix fix_data_dir.sh bug introduced in #2414; thanks, @Fhrozen (#2439)

5.4.139 b1be44eb2 2018-05-21 [scripts] Adding script to compute minDCF for speaker recognition (#2431)

5.4.140 1ab3df05a 2018-05-21 [egs] Madcat arabic: easier configuration of data locations; fix wer_filter. (#2440)

5.4.141 72d89cedd 2018-05-22 [egs] egs/tedlium/s5_r2 download Tedlium LMs from kaldi-asr.org (#2441)

5.4.142 182f3829e 2018-05-23 [src] Fix bug in rnnlm-sentence-probs (failure in non-debug builds) (#2445)

5.4.143 705138484 2018-05-24 [egs] Fix affix in run_tdnn_7p recipe (swbd) (#2446)

5.4.144 447e96498 2018-05-24 [egs] Add recipes for Speakers in the Wild (SITW) (#2422)

5.4.145 b1ae952fe 2018-05-24 [scripts,egs] Support averaging forward and backward RNNLMs (#2436)

5.4.146 6b94eb715 2018-05-24 [scripts] Add retry.pl, to be used when queue is flaky. (#2450)

5.4.147 e3556a075 2018-05-25 [src] Add missing 'fst::Connect() in parallel determinization; fixes issue #2448 (#2452)

5.4.148 e0311e063 2018-05-25 [scripts] Fix bug in retry.pl. (#2453)

5.4.149 9b23b17c7 2018-05-25 [egs] Fix a small bug in swbd 7p recipe (was not decoding some test sets) (#2454)

5.4.150 d6d49d09f 2018-05-26 [scripts] Apply minimum segment duration in vad_to_segments.sh (#2447)

5.4.151 7ffc9ddeb 2018-05-26 [scripts] Make retry.pl deal correctly with keyboard interrupts; make num-tries configurable. (#2456)

5.4.152 d23cad866 2018-05-29 [scripts] Allow floating-point cv_spk_percent in nnet1 scripts (#2459)

5.4.153 3f4f42567 2018-05-29 [egs] Fix TDNN example scripts to respect given train_cmd (#2460)

5.4.154 e01405181 2018-05-30 [src] Fix bug in lattice-oracle relating to wildcards in lattice output. (#2461)

5.4.155 a486fa139 2018-05-31 [scripts] Fix harmless bug in chain/get_egs.sh, thanks: teinhonglo@gmail.com (#2466)

5.4.156 9f3aa110b 2018-05-31 [scripts] Fix failure in make_bottleneck_features.sh (#2443)

5.4.157 0fb7950be 2018-05-31 [egs] Fix to paths in iban recipe (#2467)

5.4.158 fd545bbf6 2018-05-31 [egs] bug fix in egs/sitw/v2/run.sh (#2465)

5.4.159 6f5f7785e 2018-06-01 [doc] Fix I/O documentation example (#2468)

5.4.160 eef2a093a 2018-06-01 [scripts] Fix to steps/segmentation/internal/sad_to_segments.py (#2470)

5.4.161 e0e0bec3d 2018-06-01 [egs] Add TDNN+OPGRU recipe for multi-en (#2469)

5.4.162 c8db7a906 2018-06-02 [src] TopSort lattice-oracle lattice output (#2463)

5.4.163 04b1f7d66 2018-06-03 [egs,scripts] Byte-pair encoding (BPE) applied MADCAT Arabic OCR (#2434)

5.4.164 7659d8ba7 2018-06-05 [egs] Change renorm to batchnorm in BABEL recipes (#2476)

5.4.165 ce2eb3986 2018-06-05 [src,scripts] Add steps/nnet3/chain/get_phone_post.sh, to get phone posteriors from chain system (#2471)

5.4.166 b38d263cf 2018-06-06 [egs,scripts] Fix train/test data split in make_voxceleb1.pl, make speakers disjoint (#2475)

5.4.167 28e4ee6aa 2018-06-06 [scripts] Fix typo in validate_data_dir.sh causing failure if reco2dur is present(#2478)

5.4.168 5a6477bbc 2018-06-08 [scripts] Improve encoding compatibility of rnnlm dataprep (#2455)

5.4.169 e1abd8132 2018-06-08 [build,src] Upgrade OpenFst from 1.6.5->1.6.7 (#2313)

5.4.170 40757c2f2 2018-06-09 [scripts,build] ban carriage-return (\r) in lang/; suppress 'ar' warning (#2485)

5.4.171 0f954c272 2018-06-09 [build] Update supported compilers for CUDA 9.2. (#2486)

5.4.172 8fae9254f 2018-06-10 [build] Fix syntax error in configure script introduced in #2486 (#2488)

5.4.173 4571f47f8 2018-06-10 [scripts] Small bugfix in steps/cleanup/decode_segmentation.sh (#2489)

5.4.174 707baa0c3 2018-06-11 [egs] updating results in voxceleb recipe (#2494)

5.4.175 8fd2ba557 2018-06-11 [src] Make it so very wide-context nnet3 models don't cause a crash. (#2487)

5.4.176 be96725a7 2018-06-11 [scripts] Add script to limit <unk> history of ARPA files, for OOV-modeling efficiency (#2483)

5.4.177 1e4957873 2018-06-13 [build] change the build status badge location (#2497)

5.4.178 52b28821d 2018-06-13 [src] Fix mistake in comment RE FFT dimensions (#2495)

5.4.179 ca239f22d 2018-06-13 [scripts] Make inline script in extract_ivectors.sh pyton3 compatible (#2498)

5.4.180 b4eda5725 2018-06-14 [scripts] Make inline script in extract_ivectors.sh python3 compatible, 2nd fix (#2503)

5.4.181 f49b3fc6d 2018-06-14 [scripts,src] added a script to scale arcs which output <unk> in HCLG.fst; other small fixes (#2499)

5.4.182 bfd0d3c11 2018-06-15 [src] Fix compilation issue with DOUBLE_PRECISION flag (#2496) (#2505)

5.4.183 9210f3a78 2018-06-15 [src] Speed improvement in pruned lattice-rescoring algorithm (#2504)

5.4.184 d2b1c55ea 2018-06-15 [scripts] Make soft link from steps/chain to steps/nnet3/chain (#2507)

5.4.185 2fa70e3a6 2018-06-15 [scripts,egs] Minor updates to WSJ-e2e recipes and scripts (#2506)

5.4.186 f86792e51 2018-06-18 [src] Remove unused binary 'extract-rows' (#2510)

5.4.187 598b177c0 2018-06-19 [scripts] Bug fix to failure in chain/train.py when --trainer.optimization.do-final-combination=false (#2513)

5.4.188 775c770a1 2018-06-19 [src] Minor optimizations in "e2e" numerator code (#2508)

5.4.189 42a673a5e 2018-06-20 [egs] Add recipe for AIShell2 (large Mandarin database) (#2398)

5.4.190 d16ef4bb6 2018-06-20 [src] Change smart ptr copy to move sematics, avoid cost of atomic incr/decr (#2514)

5.4.191 60141df48 2018-06-25 [scripts] enable --trainer.input-model option for non-chain nnet3 training (#2515)

5.4.192 8ce3a9576 2018-06-27 [src] Make RNNLM training single threaded (workaround for CuSparse library bugs) (#2520)

5.4.193 e4908f079 2018-06-29 [build] disable -march=native in travis build (#2511)

5.4.194 7605fe550 2018-06-29 [src] lattice-copy: fix to segfault with --include,--exclude options if not sorted (#2412)

5.4.195 c1d671952 2018-07-01 [scripts] fixing bug present since #2515 in steps/nnet3/train_raw_{dnn,rnn}.py (#2527)

5.4.196 a34aefd6f 2018-07-02 [build] Fix typo in comment in src/configure (#2529)

5.4.197 8a27e86b4 2018-07-02 [src] Add a check for utterance length when creating end2end egs (#2528)

5.4.198 be7c1aacb 2018-07-03 [scripts] Output the reco2dur file in speed-perturbation script (#2531)

5.4.199 ad93210ad 2018-07-05 [egs] Fix to training list in egs/sitw{v1,v2} recipe (#2535)

5.4.200 6edf1be70 2018-07-06 [src] Fix to nnet3-chain-get-egs corner-case bug discovered by Armando Muscarellio (#2536)

5.4.201 998a4d657 2018-07-06 [egs] Some fixes and cleanup in Aishell2 scripts (#2522)

5.4.202 eacf34a85 2018-07-07 [src,scripts,egs] Updated TDNN-F recipes incorporating resnet skip connections. (#2430)

5.4.203 295877c2e 2018-07-07 [scripts] Add decoding script that combines nnet3 outputs (#2534)

5.4.204 00bc59df7 2018-07-09 [egs] added evaluation scripts to chime-5 setup (#2521)

5.4.205 147f0b552 2018-07-10 [egs] Make validate_text.pl say more specifically which lines have banned UTF.

5.4.206 09d670632 2018-07-10 [egs] Fix to Swbd script to reflect what I actually ran

5.4.207 ae4292dcf 2018-07-10 [egs] Fix bug in Chime5 eval script (#2542)

5.4.208 6f2140b03 2018-07-10 [scripts] Fix to get_reco2dur.sh (thanks: Zhangfeifan) (#2544)

5.4.209 fdb677438 2018-07-12 [egs] Add scripts for release 3 of Tedlium corpus (#2442)

5.4.210 2c1b11aab 2018-07-15 [src] Change decodable code so nnet context does not have to be recomputed. (#2549)

5.4.211 bbba5ef57 2018-07-15 [scripts] Add missing quote in vad_to_segments.sh script (#2550)

5.4.212 79883f30b 2018-07-16 [scripts] add utility script compute_sentence_probs_arpa.py to compute probs given arpa (#2538)

5.4.213 654d66b50 2018-07-17 [scripts] BSD-compatibilty fix where the "-s" option of seq is used (#2546)

5.4.214 c02977529 2018-07-18 [src] Speed up AddDiagMat2 for very thin and tall matrices (#2555)

5.4.215 a630dd292 2018-07-21 [egs] fixed a minor issue (wrong graph dir) in CommonVoice setup (#2563)

5.4.216 1d61f1cf4 2018-07-22 [src,scripts,egs] Misc small fixes, including fix to TDNN-F results in mini_librispeech (#2565)

5.4.217 8ae505ed8 2018-07-22 [scripts] Fix to scripts/rnnlm/get_num_splits.sh RE large datasets (#2567)

5.4.218 e6fe7e816 2018-07-24 [scripts] Removing has_key() for python3 compatibility (see #2570) (#2571)

5.4.219 de03a7525 2018-07-25 [scripts] Fix an issue in BPE-related scripts (don't assume opt-sil is phone 1) (#2547)

5.4.220 06604a8b8 2018-07-25 [src] Updated cuda kernel to speedup AddDiagMatMat[kTrans, kNoTrans] (#2560)

5.4.221 1616653d6 2018-07-25 [scripts] Fix to analyze_alignments/lats.sh (double-counting) (#2573)

5.4.222 97b78df49 2018-07-25 [egs] Update TDNN-F script for Librispeech (#2574)

5.4.223 5391f0038 2018-07-26 [src] Minor code-style fixes to context-dependency code (#2576)

5.4.224 d4d968c75 2018-07-26 [src] Add binary fsts-concat to concatenate Kaldi archives of FSTs (like fsts-union) (#2562)

5.4.225 b8fd2cd9d 2018-07-27 [egs] Set the acoustic scale properly when generating e2e chain alignments for OCR tasks (#2579)

5.4.226 ec71be48e 2018-07-27 [scripts] Fix typo in steps/cleanup/decode_fmllr_segmentation.sh (#2580)

5.4.227 5cc9731af 2018-07-27 [src] added configure support for cuda on arm64 (#2577) (#2578)

5.4.228 5b27111ae 2018-07-28 [egs] Cosmetic fix in aishell recipes (#2582)

5.4.229 68c926ab9 2018-07-31 [scripts] Call the right script when fixing/validating data directories in OCR … (#2587)

5.4.230 04027ee5d 2018-08-02 [src] fix embedding training bug for RNNLM without letter features (#2586)

5.4.231 287e24955 2018-08-04 [src] Make vector-compute-plda check that num-ivectors > plda-dim, and die if not (#2591)

5.4.232 532f3845a 2018-08-06 [build] Update check_dependencies.sh so lowest required GCC is 4.8.3 (#2596)

5.4.233 d10540cca 2018-08-07 [scripts] Fix to script usage message (#2601)

5.4.234 fec68b200 2018-08-07 [src] Correct usage message of acc-lda (#2598)

5.4.235 bee102286 2018-08-07 [scripts] RNNLM script fix: to accept successive spaces in configs (etc.) #2595 (#2597)

5.4.236 f40fa5c24 2018-08-07 [scripts] Slight cleanup in lmrescore_rnnlm_lat.sh (#2554)

5.4.237 8e97639b7 2018-08-07 [src] Fix Windows out-of-range iterator issue for nnet3 (#2594)

5.4.238 b4a8e9b82 2018-08-08 [src] Update Windows installation instructions (#2607)

5.4.239 a932be6e6 2018-08-09 [egs] Fix to LibriSpeech download script [affects 2nd run] (#2611)

5.4.240 c60f212c6 2018-08-09 [src] Change RNNLM test program to clean up temporary file (#2610)

5.4.241 6926b60a0 2018-08-11 [egs] Add a BPE-based recipe for IAM handwriting recognition (#2600)

5.4.242 94e561a1b 2018-08-11 [egs] Fix TEDLIUM v3 data download (#2609)

5.4.243 d60404b25 2018-08-11 [egs] Update LM in heroico recipe; fix bug in utils/lang/limit_arpa_unk_history.py (#2613)

5.4.244 7aa9da58d 2018-08-13 [src] Fix bug in PLDA scoring for diarization.. crashed or produced garbage if PCA failed. (#2590)

5.4.245 8c0e3e311 2018-08-13 [src] Refactor CUDA allocator code based on large cached regions (#2593)

5.4.246 cd27a83f1 2018-08-14 [build] Add missing dependencies to Dockerfiles (#2622)

5.4.247 2148d09f3 2018-08-14 [scripts] Usage-message fix in RNNLM tools (#2623)

5.4.248 e31776f8e 2018-08-16 [scripts] python3 compatibility fix to log_parse.py (#2626)

5.4.249 bba22b584 2018-08-17 [scripts] python3 compatibility fixes in nnet3/chain scripts (#2629)

5.4.250 c99932903 2018-08-17 [scripts] subset_data_dir.sh: preserve the comments in STM file (#2627)

5.4.251 094d22746 2018-08-18 [egs] adding missing callhome data-prep script; thx: Dharmesh Agrawal (#2631)

5.4.252 370ed555e 2018-08-20 [scripts] Fix broken pipe problem in get_egs_targets.sh, thx:@iezhanqingran (#2634)

5.4.253 34df7e20d 2018-08-21 [scripts] Another fix for broken-pipe issue in steps/nnet3/get_egs_targets.sh (#2635)

5.4.254 b55c36960 2018-08-23 [scripts] Add basis-fMLLR version of align_lats.sh

5.4.255 d8a686d7c 2018-08-23 [scripts] Add basis-fMLLR version of align_fmllr_lats.sh, for when there is no spk info (#2639)

5.4.256 ecb97273d 2018-08-23 [scripts] Small fix to retry.pl, RE array jobs. (#2640)

5.4.257 a56f9b34c 2018-08-23 [scripts] Compatibility fixes to basis-fMLLR scripts (write transforms per speaker) (#2642)

5.4.258 0642cb9e6 2018-08-23 [src,scripts,egs] Add CNN+TDNN example scripts (#2643)

5.4.259 ed74857da 2018-08-24 [scripts,src] Add cleanup/segmentation scripts based on nnet3 (#2581)

5.4.260 b1272d934 2018-08-24 [egs] Add CNN+TDNN-F example script for Swbd (#2644)

5.4.261 74029707e 2018-08-25 [egs] Madcat Chinese OCR recipe (#2312)

5.4.262 aa4862050 2018-08-25 [src] Cosmetic fix to gmm-mixup.cc, thx: lizhixuan635760758@gmail.com (#2645)

5.4.263 bd5111dd9 2018-08-27 [scripts] Fix bug RE segment end padding in clean_and_segment_data*.sh (#2651)

5.4.264 f788b9174 2018-08-28 [scripts] RNNLM fix: change encoding from latin-1 to utf-8 for choose_features.py (#2652)

5.4.265 f4e4b7713 2018-08-29 [src] Fix unit-test for CuMatrix, avoid occasional failure (#2654)

5.4.266 aa7568917 2018-08-29 [src] Set -O1 by default; change KALDI_ERR to suppress spurious warnings (#2411)

5.4.267 2816821f5 2018-08-30 [src] Remove pre-C++11 compatibility for noexcept. (#2656)

5.4.268 1242305c6 2018-08-30 [src] Decoder speedup: check NumInputEpsilons() in ProcessNonemitting() (#2641)

5.4.269 66145ea9e 2018-08-31 [egs] Zeroth-Korean: Korean open-source corpus and its script (#2296)

5.4.270 b842c776d 2018-09-01 [src] Clarify feature extraction options; make min-energy independent of BaseFloat (#2667)

5.4.271 e50bd7308 2018-09-01 [src] Fix typos from PR 2667 (#2669)

Version 5.5

Version 5.5 is the current master branch. The change that was made between the end of 5.4 and the start of 5.5 is support for Support for grammars and graphs with on-the-fly parts. grammar decoding; this allows support for things like the "contact list scenario" where you want to use a dynamically changing contact list in a larger, fixed decoding graph.

Below are commits corresponding to minor version numbers 5.5.x.

5.5.0 1242305c6 2018-08-30 [src] Decoder speedup: check NumInputEpsilons() in ProcessNonemitting() (#2641)

5.5.1 66145ea9e 2018-08-31 [egs] Zeroth-Korean: Korean open-source corpus and its script (#2296)

5.5.2 b842c776d 2018-09-01 [src] Clarify feature extraction options; make min-energy independent of BaseFloat (#2667)

5.5.3 e50bd7308 2018-09-01 [src] Fix typos from PR 2667 (#2669)

5.5.4 1180e467c 2018-09-02 [src] Remove `ExtractWaveformRemainder` (#2672) (#2673)