VoxForge
VoxForge
Julius 4.0 is now available from the Julius Dev site - from the release notes:
What's new in Julius-4.0
==========================
The Julius rev.4.0 is a full major version up, a re-innovation of the
decoder as a flexible speech recognition engine. The internal
structures are re-organized and modularized thoroughly, which results
in the librarization of core engine, enhancement and unification of
language model, realization of multi-decoding with hot plugging
feature. The major features are listed below:
- Engine core becomes separate C library with simple API
- Can handle various LM, thus Julius and Julian are integrated
- Multi-decoding with multiple models
- Support to add and remove models while running the decoder
- Support N-gram longer than 4 (N now unlimited)
- User-defined LM function
- Confusion network output
- GMM-based and decoder-based VAD
- New tools added, new functions added
- Memory efficiency is improved
Julius-4.0 ensures the compatibility with Julius-3.x for its usage and
Jconf configuration, so that one can easily migrate to Julius-4.
The decoding performance of Julius-4.0 is still kept as the same as
the latest release (Julius-3.5.3) for now.
Many thanks to Ralf Herzog for picking this up and posting on his blog.
Ken
--- (Edited on 12/26/2007 4:33 pm [GMT-0500] by kmaclean) ---
Documentation links:
Code Links:
- Source tarball
- Julius: julius-4.0.tar.gz (1.0MB)
- HTK-to-Julius grammar converter: slg2dfa-1.0.tar.gz
- Pre-compiled binaries
- Linux: julius-4.0-linuxbin.tar.gz (3.2MB, including PDF documents)
- Win32: julius-4.0-win32bin.zip(4.0MB, including PDF documents)
--- (Edited on 1/22/2008 4:45 pm [GMT-0500] by kmaclean) ---
From the Julius Announcements Forum page:
Julius rev.4.0.1 has been released today.
This release incorpolates bug fixes from 4.0 that has been reported so far.
http://sourceforge.jp/projects/julius/f ... 9930#29930
For Linux users, it should be noted that the default audio API
has been changed from OSS to ALSA. You can still use
the old OSS interface by giving "--with-mictype=oss" for
configure script at compilation time.
Have fun!
From the release notes:
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.
--- (Edited on 3/12/2008 3:40 pm [GMT-0400] by kmaclean) ---
--- (Edited on 3/12/2008 3:41 pm [GMT-0400] by kmaclean) ---
Hi kmaclean,
thanks for the nice information~
one question I would like to ask is: whether the user can use Confusion Network library/binary independently for other search purposes with proper inputs, e.g. from HTK enginer?
thanks!
--- (Edited on 10/3/2013 12:26 pm [GMT-0500] by xffmqjx) ---