-
Notifications
You must be signed in to change notification settings - Fork 25
Expand file tree
/
Copy pathChangeLog
More file actions
72 lines (58 loc) · 2.08 KB
/
Copy pathChangeLog
File metadata and controls
72 lines (58 loc) · 2.08 KB
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
1.0
Use shorter version number
Add JNI finder in CMake
Suppress cmd.log file
Fixes for threaded decoder from Le Zhang
Transducer loading bug-fix
Default to static libraries
0.12.1
Fix the default backoff problem in the WFST generation
0.12.0
Update for component commit in tracter
CMake build system
0.11.0
After the RT09 evaluations
o Faster WFSTDecoderLite core
o Experimental threading core
o IPP support
o Output probability look-ahead (good with IPP)
0.10.0
Lock the (internal) version for RT09 evaluations.
0.9.0
Many performance enhancements.
HTKFlatModels for faster access.
Fix decoder behaviour for tee models.
Add more phone and word pruning options.
Remove some unused code.
0.8.1
Support re-enabled for HTK extended filenames.
DecodingHMM and some other files removed
Pure virtual interface for models
0.8.0
Feature mechanism changed to use tracter.
SWIG interface to Java
0.6.1
This release is just cosmetic changes to the build environment:
o Move src/index.h to doc/index.doc
o Doxygen installs to top level, not in doc/
o README, INSTALL etc are there
o config.h.in is gone - build it with autoheader
o Header files added to the distrubution
0.6.0
gramgen now puts auxiliary symbols (#phi) into backoff loops
lexgen puts the #phi onto a loop at the initial state (with
-addPhiLoop) cdgen passes through #phi symbols.
Juicer can now take both language model match factor and insertion
penalty at the decoder level (-lmScaleFactor and -insPenalty).
The short pause transition is now handled in the decoder meaning
that the FST does not need to handle it (i.e. do not use
-pauseTeeTransProb).
lexgen can now normalise pronunciations such that the FST is
stochastic (with -normalise).
Build scripts are now sensitive to auxiliary symbols. Not using
these symbols will produce an FST that is non-determinisable, and
the tools may run out of memory before realising that it is
non-determinisable.
gramgen now has a -normalise option to make the resulting (ngram)
FST stochastic. It is not implemented efficiently, and doesn't
seem to have a noticeable effect.