1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
|
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.
|
