mirror of
https://github.com/cmusphinx/sphinxtrain.git
synced 2026-05-17 13:10:52 +00:00
David Huggins-Daines
136 lines
4.6 KiB
Plaintext
136 lines
4.6 KiB
Plaintext
SphinxTrain 5.0.0
|
|
^^^^^^^^^^^^^^^^^
|
|
|
|
New features:
|
|
|
|
* Build system now uses CMake, supports VSCode
|
|
* SphinxBase is not a requirement
|
|
* Runs in-place properly
|
|
* Docker support to ease dependency pain
|
|
* Include force-aligner from Sphinx3
|
|
|
|
Fixes:
|
|
|
|
* Feature extraction doesn't remove data from the input
|
|
* Improved error reporting
|
|
* Parallel training with Queue::POSIX is now reliable (tested on 64 CPUs)
|
|
|
|
SphinxTrain 1.0.8
|
|
^^^^^^^^^^^^^^^^^
|
|
|
|
New features:
|
|
|
|
* New openfst-based G2P trainer and decoder, supported by Sphinx4 too.
|
|
* Parallel feature extraction
|
|
* Package can be installed now just like any application
|
|
* Single 'sphinxtrain' command to access all training process
|
|
* Increased reuse of sphinxbase functions
|
|
|
|
Fixes:
|
|
* Memory leaks and build issues
|
|
|
|
Thanks: Lionel Koenig, Peter Grasch, John Salatas, Alexandru-Dan Tomescu, Yuri Orlov,
|
|
Michal Krajnansky, Jidong Tao, Tanel Alumae, Vassil Panayotov, Vijay Abharadwaj,
|
|
Melmahdy
|
|
|
|
SphinxTrain 1.0.7
|
|
^^^^^^^^^^^^^^^^^
|
|
That was two years, time to release this stuff
|
|
|
|
New features:
|
|
|
|
* MLLR !!!!
|
|
* PTM !!!!!!!
|
|
* MMIE !!!!!!!!!
|
|
|
|
* Sphinxbase is now a requirement
|
|
|
|
Improvements:
|
|
|
|
* Built-in align
|
|
* Many useful python modules
|
|
* Simple python module for reading/writing pronunciation dictionaries
|
|
* Add support for creating a single pronunciation dictionary in force alignment
|
|
* Lattice rescoring, with and without FSTs, which may be useful, plus a
|
|
little script to generate composed class language model FSTs.
|
|
* Add support for building and evaluating FSTs from class-based language models
|
|
* Save falignout files in list directory to avoid having to re-run falign in future training
|
|
* Dictionaries are assumed to be in utf-8
|
|
* Add an option to ignore word position in mk_mdef_gen
|
|
* Improve a lot of log messages
|
|
|
|
Bug fixes:
|
|
|
|
* Protect scripts from doing dangerous things
|
|
* Restrict memory usage on norm on stage 30
|
|
* Finally get around to fixing the annoying problem where the
|
|
convergence criterion differs based on the dynamic range of the
|
|
acoustic features (yes, really). It's still a ratio, just not a ratio
|
|
of logarithmic values, which didn't make much sense anyway.
|
|
* Be tolerant about whitespaces!
|
|
* Simplified log_diag_eval for gcc vectorizer
|
|
|
|
Thanks: David Huggins-Daines, Long Qin, Tanel Alumae, Stephan Vanni, and everyone else!
|
|
|
|
SphinxTrain 1.0
|
|
^^^^^^^^^^^^^^^
|
|
|
|
This release corresponds to revision 8978.
|
|
|
|
Since there has never been an official release of SphinxTrain, we felt
|
|
it was necessary to create one to point people to and for them to
|
|
report bugs against, instead of the confusion of having to deal with
|
|
whatever was currently in SVN.
|
|
|
|
Compared to past unofficial releases, this version is significantly
|
|
improved. There is support for speaker adaptation, feature space
|
|
transforms (LDA and MLLT), and it also includes "templates" which can
|
|
be used to train models from standard corpora from the LDC.
|
|
|
|
Sphinxtrain pre-1.0
|
|
^^^^^^^^^^^^^^^^^^^
|
|
|
|
From 2000, when SphinxTrain is first opened to the public, it was
|
|
regarded as merely an auxillary tool for the decoders (such as Sphinx
|
|
2, Sphinx3 and later Sphinx 4). However, many users in the forum
|
|
asked about how training could be done and most of the time they would
|
|
touch a lot of bugs in the software. There is a great need to
|
|
continue the support of SphinxTrain and increase its quality.
|
|
|
|
We hope that this release will
|
|
|
|
* fix significant amount of bugs of the code and the scripts.
|
|
* improve the quality of the code and scripts
|
|
* extend the functionalities of SphinxTrain.
|
|
* ensure SphinxTrain is fully regression tested.
|
|
|
|
New Features:
|
|
|
|
* mllr_solve, could be used to find the regression matrix
|
|
for MLLR (r3566, 3567, 3731)
|
|
* Added the option "-input_endian" so that one can specify
|
|
byte. It follows the the compilation definition (r498, r714)
|
|
order of raw data as different from the machine native byte order.
|
|
* Prints out links to log files in the html file generated during
|
|
training, and creates possibly missing directories (r580)
|
|
|
|
Bugs:
|
|
|
|
* norm was requesting some parameters that it doesn't really need,
|
|
like feature string definition and size of input vector. Removed the
|
|
request.
|
|
* add checking for wave2feat
|
|
* When transcription is empty, corpus.c wouldn't show abnormal
|
|
behaviour now
|
|
* Memory problems in wave2feat.
|
|
|
|
General Code Architecture Related:
|
|
|
|
* Initial checkin by Prof. Alan Black. That includes
|
|
basic Makefiles, RickyH's perl scripts.
|
|
* Miscellaneous fixes of Ricky's perl script.
|
|
|
|
Developers:
|
|
|
|
* check for return values in read/write.
|