Release.txt

4.6 (2020.09.02)
=================
New features:
- CUDA support for DNN computation
  - build by nvcc to enable it.
  - detailed parameter can be given by .dnnconf option "cuda_mode=...".  See Sample.dnnconf for details.
  - tested on Linux with CUDA 8.0, 9.0 and 10.2.
- 1-pass grammar recognition
  - Per-grammar basis: can be enabled for each grammar only when it has additional ".dfa.forward" file.
  - The ".dfa.forward" file will be generated by "mkdfa" of recent version.  Leave it to enable, or delete it to let Julius work as previous version.
- Support non-log10nized state priors in DNN model
  - New .dnnconf option "state_prior_log10nize=yes/no" to switch the behavior
- Feature normalization pattern added: mean = input self, variance = static
  - New option "-cvnstatic" to choose this behavior
  - See the updated doc "doc/Normalize.md" to know how to set feature normalization in Julius.
Updates:
- Now delivered under simplified BSD License
- Added Python version of "mkdfa.py"
- Update build for Visual Studio 2017, support building more tools.
- Re-write documentations in markdown format under "doc" (WIP)
- Place README.md in each directory, remove *.txt instead
- mkdfa (mkfa): now outputs detailed error message (line num etc.) 
Bug fix
- "mkbingram" ignores charset conversion options, performs no conversion


4.5 (2019.01.02)
=================
New features:
- Improve voice detection by integrating "libfvad", a voice activity
  detection library based on WebRTC's VAD engine  [https://github.com/dpirch/libfvad]
  - Now Julius has dual-mode VAD:
     - old module (input level and zero-cross based)
     - new module (libfvad = model based)
  - both runs in parallel: both modules runs for an audio input stream
    concurrently
  - detect speech only when **both module triggers**!
  - new module is disabled by default
     - apply "-fvad arg" option to enable
     - arg is a switch, "-fvad 0" for moderate mode and "-fvad 3" for aggressive mode.
  - new module is available on all audio modules
     - julius
     - adinrec
     - adintool
     - adintool-gui
  - typical usage:
      - "-fvad -1" to use old VAD only (same as older versions)
      - "-fvad 3 -lv 1" to use new VAD only.  "-lv 1" forces the old VAD
      	to "always triggering", thus the final VAD result fully depends on
	the new module.
- New multi-threaded DNN computation
   - added "num_threads" option to dnnconf to specify number of CPU
   threads to be used on DNN computation.
   - default number of thread to be used is 2.

Modified:
- module output now performs XML Escape
  - characters <, >, ", & and ' in output string are now escaped to
    &lt;, &gt;, &quot;, &amp; and &apos;.  This escaping is enabled by
    default from this version, however you can switch off this
    escaping and keep old behavior by applying "-noxmlescape" option.

Bug fixes:
- fix Makefile for parallel build (make -j N)
- fix adintool-gui sometimes segfault with arguments
- fix several build warnings
- fix several memory leaks
- fix mis-compilation on some OS

[New run-time options]
  [-fvad mode]  set libfvad mode. "mode" is an integer value from -1 to 3, -1 to disable,
  0 for moderate detection, 3 for aggressive detection (more likely to
  drop speech-like noises).  Default value is -1 (disabled)

  [-fvad_param nFrame thres]  set libfvad detailed parameter.  "nFrame"
  is the number of smoothing frame. "thres" is the threshold to
  detect speech trigger [0.0-1.0].  Default values are 5 and 0.5
  respectively.

[New configure options]
- "--disable-libfvad": disable libfvad integration

[Document update in progress]
- adinrec,adintool README.md README.ja.md
- julius README.md, Options.md

4.4.2.1 (2016.12.20)
====================
- Small fixes for Android and iOS.
- Clean up msvc dir.

4.4.2 (2016.09.12)
===================
- Improved handling of file paths in dnnconf, now correctly handled as relative to the dnnconf file.
- Improved DNN decoding that sometimes goes too slow and stack on 2nd pass.
- Fix segfault on old non-AVX Intel CPU with DNN.
- Fix errors in build process on ARM and VisualStudio.

4.4.1 (2016.09.07)
===================
- more stable and fast SIMD code: SSE, FMA and ARM_NEON
- automatically select suitable SIMD code at run time for DNN computation
- msvc support updated: PortAudio and zlib sources are now included in dist.
- fix incorrect reading of binary hmmlist made by "mkbinhmmlist"
- fix SDL detection in adintool-gui
- "INSTALL.txt" to share how to build Julius on various platform.
- pkg-config support
- other fixes

4.4 (2016.08.30)
=================
- DNN-HMM computation support
- "adintool-gui": adintool with input monitoring (see adintool/README-GUI.txt)
- "binlm2arpa": convert binary LM to ARPA format
- "mkbingram" now can convert text encoding of an LM by "-c" option
- fix not to exit at disconnection on module mode, wait for next instead.
- fix compilation errors in some recent OS
- fix memory leaks
- work on autoconf >=2.6
- added README.md, CONTRIBUTING.md and other files for GitHub hosting
- added document to use Julius with DNN-HMM AM: "00readme-DNN.txt"
- update support for VS2013

4.3.1 (2014.01.15)
===================
Fixed bugs:
- Compilation error on OS X.
- Unnecessary debug messages in adintool.
- Several bugs around reading / applying "-cmnload".

4.3 (2013.12.25)
=================
New features:
- FBANK and MELSPEC support.
- Network-based feature vector and outprob vector input.
- Static mean/variance for cepstral mean/variance normalization.
- State output probability (i.e. outprob) vector input for DNN-HMM decoding.
- State ID "<SID>" extension of hmmdefs for DNN-HMM decoding.
- Real-time feature extraction and network transmittion by 'adintool'.

Modified:
- "mkbinhmm" now keeps the state order and id of the original hmmdefs.
- For portaudio, pause / resume operation synced between engine and audio I/O
- Load / save cepstral mean/variance of CMN/CVN in HTK text format.

New options:
  [-input vecnet]       read feature / outprob vectors from network
  [-input outprob]      read outprob vectors from HTK parameter file
  [-outprobout [file]]  save computed outprob vectors to HTK file (for debug)


4.2.3 (2013.06.30)
==================

New features:
- Add function "j_reload_adddict()" to reload dictionaries.
- Add option "-lvscale factor" and func "j_adin_change_input_scaling_factor()"
  to scale the amplitude of captured audio by the factor.
- Add option "-rejectlong msec" to reject too long input.
- Add minimum bayes risk decoding, contributed by H. Nanjo and R. Furutani
- Support binary N-gram symbol charset conversion by "mkbingram".

Fixes:
- Fix sending audio stream via network with incorrect byte order at
  big-endian machines.
- Fix occasional failure of closing audio device at j_close_stream().
- Fix segfault when reading binary hmm created at 64bit env. with embedded parameters.
- Fix memory leak when failed to read an N-gram file.
- Fix memory leak when input length overflow is detected.
- Fix unable to load feature vector plugin.
- Update microphone input code for recent MacOSX.


4.2.2 (2012.08.01)
==================

Fixes:
- Now can be compiled without flex library
- Fix failure of reading binary N-gram when compiled with "--enable-words-int"
- Fix incorrect handling of file paths with backslash in jconf file at Windows
- Fix segfault when reading an errorous word dictionary.
- Fix occasional segfault which may occur while search.


4.2.1 (2011.12.25)
===================

New features:
- Add support for per-word insertion penalty setting at grammar
  recognition. You can set different word insertion score for each word
  entry at .dict file. For example, if you have an entry
        15 [a] a
  in .dict file and want to assign word insertion score of "-2.0" to
  this word, you can write like this:
        15 @-2.0 15 [a] a
  The figure after "@" is the insertion penalty. The third
  element should be the same as the first element.

- New option "-chunk_size" can specify the audio fragment size in
  number of samples. The default value is 1000.

- At "adintool", enable input detection by default for standard input.

Fixed bugs:
- (IMPORTANT) CMN is not performed for C0 coef.  This bug exists in
  the versions from 4.1.3 to 4.2.
- "-forcedict" won't work for additional dictionaries given by "-adddict".
- Corrupted header of recorded WAV file when interrupted by CTRL+C.
- Occasional segfault when reading a wrongly formatted dictionary.
- Won't compile with configure option "--enable-word-graph".
- Segfault of "mkbingram" and "generate-ngram" at cygwin.


4.2 (2011.05.01)
=================

New features:
- Additional score-based pruning at the 1st pass.  It is disabled by
  default, you can enable by using an option "-bs arg". The argument
  is score range.
- New support for PulseAudio (--with-mictype=pulseaudio)
- New Option "-adddict", "-addword" to read additional dictionaries / words.
- Portaudio library updated to V19.  Audio capture device can be
  changed by env. "PORTAUDIO_DEV_NUM".  The device list will be
  output at start up.

Changed behavior:
- "mkbinhmmlist" now saves pseudo phone list extracted from AM for
  faster start up. The output should be used with the same AM
  specified at generation.  Note that the converted binhmmlist file
  can not be used with older Julius.
- Audio library linking was modified at configure script.
  When "--with-mictype=..." is explicitly specified, Julius will link
  ONLY the audio library.  If not specified, Julius will link all the
  audio devices whose development file was detected by the configure.

Library functions:
- j_config_load_string_new(char *str): like j_config_load_file(), but
  parse the given string to set parameters.
- add_dict(), add_word(): the same as "-adddict" and "-addword".
  (They should be called at start up before starting engine)
- (portaudio/Windows) j_open_stream(recog, NUMSTR) to choose device NUM.
  ex. 'j_open_stream(recog, "1")' will open device number one.
- (portaudio/Windows) get_device_list(): obtain list of available devices.

Fixes:
- Improved tree lexicon structure for better memory management.
- Reduce malloc calls at reading N-gram.
- Eliminated memory leaks using Valgrind.
- Workarounds to avoid crash with j_close_stream().
- Now allow "-iwsp" only with multi-path acoustic model.


4.1.5.1 (2010.12.25)
=====================

Modified:
 - Fixed problem related to the license.

4.1.5 (2010.06.04)
===================
Bug fixes:

 - Language model / decoding (these bugs may affect the ASR performance):
   - Several wrong word insertion penalty handling on grammar was
     found and fixed.
   - Now correctly add the prob. of the first word at the second pass.

 - MFCC computation:
   - Support MFCC computation when liftering parameter (CEPLIFTER) = 0.

 - Compilation:
   - Fixes to build Julius on cygwin and MSVC.
   - Supports "gcc -mno-cygwin" on cygwin.
   - Compilation error with configure "--disable-plugin"

 - Module mode:
   - Unable to send grammar from jcontrol.
   - Not working "DELPROCESS" command when SR and LM has different names.

 - Other fixed bugs:
   - wrong parsing of "-mapunk" option.
   - "-htkconf" in a jconf file now correctly handles the file path as
     relative to the jconf file.
   - "-input stdin" now supports WAV format.
   - not working "-plugin DIRNAME" on Win32/MSVC.


4.1.4 (2009.12.25)
===================

New feature:
 - added function to choose input audio device on MSVC compiled Julius,
   by specifying a device ID with env. var. "PORTAUDIO_DEV_NUM".
   The available device IDs will be listed in the system log at start up.
 - You can now set a locale for a LM in Julius.cpp.

Bug fixes:
 - now can be compield on Mac OS X (OS X 10.6 SDK).
 - fixes around portaudio for smaller latency and compatibility (Windows).


4.1.3 (2009.11.02)
===================

New features:
 - new MSVC support: please read "msvc/00README.txt"
 - extended N-gram to support arbitrary N
 - portaudio external library (V19) can be used instead of internal V18.
   When configure detects portaudio library installed in your system,
   Julius will use it instead of internal V18.  You can also choose
   input device by "PORTAUDIO_DEV" env. var. at V19library.  See the
   log text at start up to know how to set it.
 - allow word alignment output (-walign) in module mode

Modified:
 - ! now Julius do not perform CMN on 0'th cepstral coefficients,
   which is the same as the old 4.0.x versions.
 - j_get_current_filename() added on JuliusLib
 - improved "--enable-wpair" handling

Bug fixes:
 - many bugs around audio open/close API on JuliusLib
 - fail to do make in julius-simple
 - unable to record inputs at cygwin
 - segfault on adintool with "-server"
 - occasional segfault at grammar recognition


4.1.2 (2009.02.12)
===================

[SRILM support]
 - Added swapping "<s>" and "</s>" when reading BACKWARD ARPA file
   trained by SRILM.  It will be automatically detected.  If detection
   fails, you can specify an option "-swap" in mkbingram to do that.
 - Internally modify the unigram probability of "<s>" or "</s>", since
   they may be set to "-99" in SRILM model.  The same value as
   opposite will be assigned.
[N-gram]
 - Size limit extended from 2GB to 4GB for big N-gram.
 - "<unk>" and "<UNK>" can be changed by "-mapunk".
 - More strict check for unknown words: Julius now terminates with
   error when dictionary has OOV words and N-gram is not open (no unk word).
[Improvements]
 - Faster successor list building algorithm
 - Update yomi2voca.pl to cover more minor Japanese pronunciation.
 - Workaround for audio buffer overrun in ALSA
[JuliusLib]
 - Added API function "j_close_stream()" to exit main recognition loop.
[Bug Fixes]
 - Fixed segfault on adintool when specifying multiple servers.
 - Fixed compilation error on cygwin (libesd)
 - Fixed segfault when not specifying "-input" option.


4.1.1 (2008.12.13)
===================
Bug fixes:

[N-gram]
  - sometimes could not read an ARPA N-gram file trained by SRILM.
[A/D-in]
  - "-input stdin" does not work.
  - "SOURCERATE" at "-htkconf" is ignored.
[Forced alignments]
  - now can be used in isolated word recognition and with "-1pass".
  - "-palign", "-walign" and "-salign" can not be run together at a time.
[Module mode]
  - freezes when a grammar is specified by its ID number.
  - wrong grammar ID in recognition result (GRAM=.. always 0)
  - "SYNCGRAM" will cause crash at isolated word recognition.
  - unable to receive/activate/dactivate on isolated word recognition.
[Others]
  - fails to compile on several OS (needs "-ldl").
  - does not handle backslash escaping correctly in Jconf file.
  - does not output the 1st pass result as a final result with "-1pass".
[Tools]
  Jcontrol
  - does not support "graminfo" command.
  - can not send a dictionary to Julius running isolated word recognition.
  mkdfa
  - segfault on mkfa
  - fails to read a grammar file on DOS format.
  adintool
  - wrong behavior when splitting a long audio file.
  - now output time of each segment.


4.1 (2008.10.3)
================
New plugin extension:
  - supported types:
      - A/D-in plugin
      - feature vector input plugin
      - audio input monitor / postprocess plugin
      - feature vector monitor / postprocess plugin
      - result plugin
      - can add arbitrary JuliusLib callback via plugin
  - sample codes is included, with full documentation of function spec.
  - run on Linux, Windows and other unix variants with dlopen() capability

Newly supported features:
  - multi-stream feature input
  - MSD-HMM (compatible with "HTS" toolkit)
  - CVN
  - frequency warping for VTLN (no estimation yet)
  - "-input alsa", "-input oss" and "-input esd"
  - perl version of jcontrol client "jclient-perl"

Modified:
  - Restrict option orders when multiple instances defined (-AM, -LM, -SR):
      - Option should be just after correspondence instance declaration.
        (ex. LM options should be placed after "-LM" and before other
        instance declaration.)
      - Global option should be before any instance declaration, or
        just after "-GLOBAL" option.
    This new restriction can be removed by "-nosectioncheck" option.

Fixed bugs:
  - "-record" fails to record the first silence part!
  - Not working "-multigramout"
  - environment variable expansion sometimes fail within jconf file.
  - limits extended:
     maximum HMM name length = 256 char, Number of HMM states unlimited.
  - Module mode error message on grammar command.

Documents:
  - Alpha version of "Juliusbook" (contains only manuals at this time)
  - Unix manuals are moved to "man" directory.


4.0.2 (2008.5.27)
==================
New features:

  - New option "-fallback1pass" will output 1st pass result as final result
    when the 2nd pass fails.
  - Added support for "USEPOWER=T" on feature extraction.

Modified:
  - "-AM_GMM" becomes optional: GMM will share AM params if not specified.

Fixed:
  - GMM rejection does not work (since 4.0.1)
  - Cannot specify other A/D device on Linux/ALSA correctly.
  - Sometimes fails to read a big N-gram.
  - Sometimes crush with "-record" option.
  - Callback timing modified on real-time input with sp-segment/GMM/VAD.
  - Other minor fixes.


4.0.1 (2008.3.12)
==================
New features:

  A/D-in
    - ALSA now become default on Linux instead of OSS.

  Module mode
    - "ACTIVATEGRAM", "DEACTIVATEGRAM" and "DELGRAM" now accepts
      grammar name as arguments in addition to grammar ID number.
    - new command "GRAMINFO" to get list of current grammars.


Fixed bugs:

  A/D-in
    - ALSA codes updated to work on 1.x drivers.
    - segfault with "-48".
    - segfault on MFCC input with zero frames with "-spsegment".
  VAD
    - CMN not working on spsegment/GMM-VAD/decoder-VAD with microphone input.
  Acoustic model
    - Error when no short-pause model defined in multi-path mode.
  N-gram
    - incorrect 2-gram prob on 1st pass with backward N-gram only.
    - incorrect 1-gram prob for unknown words.
    - fail to read some ARPA files with no back-off compaction.
    - read failure or segfault on big N-gram with over 24bit entries.
    - redundant index for back-off weights in some case.
  Word recognition
    - incorrect N-best output with "-output N" on word recognition.
  Installation
    - "make install" fails on cygwin.
  Source code
    - Static variables in functions that are not meant to be static
      are made local.
    - Global variables in search are moved to StackDecode.



4.0 (2007.12.19)
=================
For more detail about new features in 4.0, please see other document.

- Re-constructed all data structures and re-organize source code.
- Core engine now becomes a library called JuliusLib, with API and callbacks.
- Multi-model decoding now available.
- Modularize language model handling, and merge Julian to JuliusLib.
- Support longer N-gram (N > 3).
- User-defined LM function support.
- Handy isolated word recognition mode.
- Confusion network output.
- Improvements in short-pause segmentation, especially for live input.
- GMM-based VAD.
- Decoder-based VAD.
- Integrated many compile-time options.
- Reduce memory usage.
- Sample application to use the JuliusLib is included: "julius-simple".
- Update tools:
   - "adintool" supports multi-server mode.
   - "generate-ngram" newly added to generate sentences from N-gram


3.5.3 (2006.12.29)
===================
  o  Improved Performance:
     - acoustic computation optimized: now becomes 20%-40% faster!
     - optimize memory access: re-use work area of deleted hypothesis
       in the 2nd pass.
     - some memory allocation improvement on dictionary and word trellis.

  o  New Grammar Tools:
     - "dfa_minimize", "dfa_determinize" will minimize/determinize DFA.
        mkdfa.pl now calls dfa_mimize in it.
     - "slf2dfa": a toolkit to convert HTK slf to Julian dfa (separate kit)

  o  Embedding HTK Acoustic Parameters:
     - add option to load HTK Config file to set correct acoustic parameter
       configuration at recognition time.
     - the acoustic parameter configuration can be embedded into
       header of a binary HMM file.

  o  Improved Word Graph:
     - add an option to completely separate graph words: words with
       different phone contexts can be output separatedly by
       "-graphrange -1".

  o  Support for online energy normalization:
     - Preliminary support for live recognition using acoustic model with
       energy normalization. (approximate with maximum energy of last input)

  o  Code refinements:
     - re-organize libsent/src/wav2mfcc.
     - modularize acoustic parameter (Value) handling.
     - output compile-time configuration of libsent with "--setting" option.
     - Doxygen 1.5.0 support.
     - "julius-info@lists.sourceforge.jp" becomes the official contact address.
     - fixed typo on copyright notice.

  o  Fixed bugs:
     - sometimes unable to read a binary LM on "--enable-words-int".
     - memory leaks around option handling, global variables and local buffers.
     - segmentation fault on very long input.
     - doublely counted initial state of DFA.
     - mkdfa.pl: unable to find mkfa on some OS.
     - adintool: makes empty output file on termination.
     - adintool: miss last inputs when killed.
     - other small changes.


3.5.2 (2006.07.31)
===================
  o  Speed-up and improvement on Windows console:
     - Support DirectSound for better input handling
     - Support input threading utilizing callback API on portaudio.
     - Support newest MinGW (tested on 5.0.2)

  o  More accurate word graph output:
     - Add option to cut the resulting graph by its depth
       (option -graphcut, and enabled by default!)
     - Set limit for post-processing loop to avoid infinite loop
       (option -graphboundloop, and set by default)
     - Refine graph generation algorithm concerning dynamic word merging
       and search termination on the second pass.

  o  Add capability to output word graph instead of trellis on 1st pass:
     - 1st pass generates word graph instead of word trellis as
       intermediate result by specifying "--enable-word-graph".
       In that case, the 2nd pass will be restricted on the graph, not
       on the whole trellis.
     - With "--enable-word-graph" and "--enable-wpair" option, the
       first pass of Julius can perform 1-pass graph generation based
       on 2-gram with basically the same algorithm as other popular
       word graph based decoders.

  o  Bug fixes:
     - configure script did not work on Solaris 8/9
     - "-gprune none" did not work on tied-mixture AM
     - Incorrect error message for AM with duration header other than "NULLD"
     - Always warns abount zero frame stripping upon MFCC

  o  Imprementation improvements:
     - bmalloc2-based AM memory management


3.5.1 (2006.03.31)
===================
  o  Wider MFCC types support:
     - Added extraction of acceleration coefficients (_A).  Now you
       can recognize waveform or microphone input with AM trained with _A.
     - Support all MFCC qualifiers (_0, _E, _N, _D, _A, _N, _Z) and their
       combination
     - Support for any vector lenth (will be guessed from AM header)
     - New option: "-accwin"
     - New option "-zmeanframe": frame-wise DC offset removal, like HTK
     - New options to specify detailed analysis parameters (see manual):
          -preemph, -fbank, -ceplif, -rawe / -norawe,
          -enormal / -noenormal, -escale, -silfloor

  o Improved microphone / network recognition by MAP-CMN:
     - New option "-cmnmapweight" to change MAP weight
     - Option "-cmnload" can be used to specify the initial cepstral
       mean at startup
     - Cepstral mean of last 5 second input is used as an initial mean
       for each input.  You can inhibit updating of the initial mean
       and keep the value loaded by "-cmnload" by option "-cmnnoupdate".

  o Module issue:
     - Julius now outputs "<STARTPROC/>" when recognition starts, and
       "<STOPPROC/>" after recognition stopped by module command.
       Use this for safer server-client synchronization.
     - now can specify grammar name from client by specifying a name
       after a command like "ADDGRAM name" or "CHANGEGRAM name".

  o Bug fixes:
     - Sometimes segfault on pause/resume command on module mode while input.
     - Can not read N-gram with tuples > 2^24.
     - Can not read HMM with 3-state (1 output state) model on multi-path.
     - Sometimes omit the last transition definition in DFA file.
     - Sometimes fails to compile the gramtools on MacOSX.


3.5 (2005.11.11)
=================
  o  New features:
     - Input verification / rejection using GMM (-gmm, -gmmnum, -gmmreject)
     - Word graph output (--enable-graphout, --enable-graphout-nbest)
     - Pruning on 2nd pass based on local posterior CM (--enable-cmthres)
     - Multiple/per-grammar recognition (-gram, -gramlist, -multigramout)
     - Can specify multiple grammars at startup: "-gram prefix1,prefix2,..."
       or "-gramlist listfile" where listfile contains list of prefixes.
     - General output character set conversion "-charconv from to"
       based on iconv (Linux) or Win32API+libjcode (Windows)

  o  Improved audio inputs on Linux:
     - ALSA-1.x support. (--with-mictype=alsa)
     - EsounD daemon input support. (--with-mictype=esd)
     - Fixed some bugs on USB audio input.
     - Audio capturing device can be specified via env. "AUDIODEV".
     - Extra microphone API support using portaudio and spLib API.

  o  Performance improvements:
     - Reduced memory size for beam operation on the 1st pass.
     - Slightly optimized tree lexicon by removing redundant data.
     - Reduced size of word N-gram index (reduced from 32 bit to 24 bit).

  o  Fixed bugs:
     - Not working spectral subtraction.
     - Memory leak when stack exhausted ("stack empty") on 2nd pass.
     - Segmentation fault on a very short input of 1 to 4 frames.
     - AM trained with no CMN cannot be used with waveform/mic input.
     - Wrong short-pause word handling on successive decoding mode.
       (--enable-sp-segment)
     - No output of "maxcodebooksize" at startup.
     - No output of the number of sentences found when stack exhausted.
     - No output of "-separatescore" on module mode.
     - Beam width does not adjusted when grammar has been changed and
       full beam options (-b 0) is specified in Julian.
     - Wrong update of category-aware cross-word triphones when
       dynamically switching grammar on Julian.
     - No output of grammar to stdout on multiple grammar mode.
     - Unable to send/receive audio data between different endian machines.
     - (Linux) crash when compiled with icc.
     - (Linux) some strange behavior on USB audio.
     - (Windows) confuse with CR/LF newline inputs in several text inputs.
     - (Windows) mkdfa.pl could not work on cygwin.
     - (Windows) sometimes fails to read a file when not using zlib.
     - (Windows) wrong file suffix when recording with "-record" (.raw->.wav)

  o  Unified source code:
     - Linux and Windows version are integrated into one source.
     - Multi-path version has been integrated with the normal version
       into one source.  The multi-path version of Julius/Julian, that
       allows any transitions of HMMs including model skip transition,
       can be compiled by "--enable-multipath" option.  The part of
       source codes for the multi-path version can be identified
       by the definition "MULTIPATH_VERSION".

  o  Other improvements:
     - Now can be compiled on MinGW/MSYS on Windows
     - Totally rewritten comments in entire source in Doxygen format.
       You can generate fully browsable source documents in English.
       Try "make doxygen" at the top directory (you need doxygen installed)
     - Install additional executables of julius/julian with version and setting
       names like "julius-3.5-fast" when "make install" is invoked.
     - Updated LICENSE.txt with English translation for reference.

  o  Changed behaviors:
     - Binary N-gram file format has been changed for smaller size.
       The old files can still be read directly by julius, in which
       case on-line conversion will be performed at startup.
       You can convert the old files (3.4.2 and earlier) to the new
       format with the new mkbingram by involing the command below:
	       "mkbingram -d oldbinary newbinary"
       Please note that since mkbingram now output the new format
       file, it can not be read by older Julius.
       The binary N-gram file version can be detected by the first 17
       bytes of the file: old format should be "julius_bingram_v3" and
       new format should be "julius_bingram_v4".
     - Byte order of audio stream via tcpip fixed to LITTLE ENDIAN.
     - Now use built-in zlib by default for compressed files.  This may
       make the engine startup slower, and if you prefer, you can still
       use the previous method using external gzip command by specifying
       "--disable-zlib".
     - (Windows) Changed the compilation procedure on VC++.  You can build
       Julian by only specifying "-DBUILD_JULIAN" at compiler option,
       and do not need to alter "julius.h".


3.4.2 (2004.03.31)
===================
- New option "-rejectshort msec" to reject short input.
- More stable PAUSE/RESUME on module mode with adinnet input.
- Bug fixes:
  - Memory leak on very short input.
  - Missing Nth result when small vocabulary is used.
  - Hang up of "generate" on small grammar.
- Cosmetic changes:
  - Cleanup codes to confirm for 'gcc -Wall'.
  - Update of config.guess and config.sub.
  - Update of copyright to 2004.

3.4.1 (2004.02.25)
===================
- AM and LM computation method is slightly modified to improve search
  stability of 2nd pass.  These modification are enabled by default, and
  MAY IMPROVE THE RECOGNITION ACCURACY as compared with older versions.
  - fixed overcounting of LM score for the expanded word.
  - new inter-word triphone approximation (-iwcd1 best #) on 1st
    pass.  This new algorithm now becomes default.
- Newly supports binary HMM (original format, not compatible with HTK).
  A tool "mkbinhmm" converts a hmmdefs(ascii) file to the binary format.
- MFCC computation becomes faster by sin/cos table lookup.
- Bugs below have been fixed:
  - (-input adinnet) recognition does not start immediately after speech
		     inputs begin when using adinnet client.
  - (-input adinnet) together with module mode, speech input cannot
		     stop by pause/terminate command.
  - (-input adinnet) unneccesary fork when connecting with adinnet client.
  - (-input rawfile) error in reading wave files created by Windows
                     sound recorder.
  - (CMN) CMN was applied any time even when acoustic models does not want.
  - (AM) numerous messages in case of missing triphone errors at startup.
  - (adintool) immediately exit after single file input.
  - (sp-segment) fixed many bugs relating short pause word and LM
  - (sp-segment) wow it works with microphone input.
  - (-[wps]align) memory leak on continuous input.
- Add option to remove DC offset from speech input (option -zmean).
- (-module) new output message:
  '<INPUTPARAM FRAMES="input_frame_length" MSEC="length_in_msec">'
- Optional feature "Search Space Visualization" is added (--enable-visualize)
- HTML documentations greatly revised in doc.

New argument: "-iwcd1 best #" "-zmean"
New configure option: "--disable-lmfix", "--enable-visualize"

3.4   (2003.10.01)
===================
- Confidence measure support
  - New parameter "-cmalpha" as smoothing coef.
  - New command "-outcode C" to output CM in module output
  - Can be disabled by configure option "--disbale-cm"
  - Can use an alternate CM algorithm by configure option "--enable-cm-nbest"
- Class N-gram support
  - Can be disabled by configure option "--disable-class-ngram"
  - Factoring basis changed from N-gram entry to dictionary word
- WAV format recording in "adinrec", "adintool" and "-record" option
- Modified output message
    startup messages,
    engine configuration message in --version and --help,
- Fixes:
    some outputs in module mode,
    bug in only several frame input (realtime-1stpass.c),
    long silence at end of segmented speech
    miscompilation with NetAudio,
    word size check in binary N-gram,
    bug in acoustic computation (gprune_none.c).
    "-version" -> "-setting", "-hipass" -> "-hifreq", "-lopass" -> "-lofreq"

3.3p4 (2003.05.06)
===================
- Fixes for audio input:
  - Fix segfault/hangup with continuous microphone input.
  - Fix client hangup when input speech too long in module mode.
    (now send an buffer overflow message to the client)
  - Fix audio input buffering for very short input (<1000 samples).
  - Fix blocking handling in tcpip adin.
- Some cosmetic changes (jcontrol, LOG_TEN, etc.)

3.3p3 (2003.01.08)
===================
- New inter-word short pause handling:
  - [Julius] New option added for short pause handling.  Specifying
    "-iwspword" adds a short-pause word entry, namely "<UNK> [sp] sp sp",
    to the dictionary.  The entry content to be changed by using "-iwspentry".
  - [multi-path] Supports inter-word context-free short pause handling.
    "-iwsp" option automatically appends a skippable short pause model at
    every word end.  The added model will also be ignored in context
    modeling.  The short pause model to be appended by "-iwsp" can be
    specified by "-spmodel" options.  See documents for details.
- Fixes for audio input:
   - Input delay improved: the initial response to mic input now
     becomes much faster than previous versions (200ms -> 50ms approx.).
   - Would not block when other process is using the audio device, but
     just output error and exit.
   - Update support for libsndfile-1.0.x.
   - Update support for ALSA-0.9.x
     (to use this, add "--with-mictype=alsa" to configure option.)

3.3p2 (2002.11.18)
===================
- [multi-path version] Supports model-skip transition.  From
  this version, you can use "any" type of state transition in HTK
  format acoustic model.
- New feature: "-record dir" records speech inputs sucessively
  into the specified directory with time-stamp file names.
- fix segfault on Solaris with "-input mfcfile".
- fix blocking command input when using module mode and adinnet together.
- modified the output flush timing to make sure the last recognition
  result will be output immediately.

3.3p1 (2002.10.15)
===================
Following bugs are fixed:
- Fixed incorrect default value of language weights for second pass (-lmp2).
- Fixed sometimes read failure of dictionary file (double space enabled).
- Fixed wrong output of "-separatescore" together with monophone model.

3.3 (2002.09.12)
==================
The updates and new features from rev.3.2 is shown below.

- New features added:
        - Server module mode - control Julius (input on/off, grammar switching)
          from other client process via network.
        - Online grammar changing and multi-grammar recognition supported.
- Noise robustness:
        - Spectral subtraction incorporated.
- Support more variety of acoustic models:
        - "multi-path version" is available that allows any transition
           including loop, skip and parallel transition.
- A little improvement of recognition performance by bug fixes
- Other minor extensions (CMN parameter saving, etc.)
- Many bug fixes

English documents are available in
  o online manuals (will be installed by default), and
  o Translated full documentation in PDF format: Julius-3.2-book-e.pdf.
We are sorry that current release contains only documents for old rev.3.2.
We are now working to update it to catch up with the current rev.3.3 version.