root / trunk / Linux / apps / codeblock / CommunityCoreAudio / bin / data / sphinxmodel / README @ 59

View | Annotate | Download (1.2 KB)

1
This is English Voxforge acoustic model for CMUSphinx decoders -
2
pocketsphinx and sphinx4 version 0.1.3. It was trained with SphinxTrain.
3
The process of training is described in SphinxTrain tutorial.
4
5
This is 8kHz model, it could be used both for desktop and telephone
6
recordings.
7
8
To setup the files use build.sh script as a base. It should download
9
required files from Voxforge, setup structure and extract features.
10
Scripts are located in scripts subfolder.
11
12
Testing is performed with sphinx3. Language model for testing is built
13
from test prompts with SRILM. Result of the testing should be similar to
14
15
TOTAL Words: 44025 Correct: 40336 Errors: 4388
16
TOTAL Percent correct = 91.62% Error = 9.97% Accuracy = 90.03%
17
TOTAL Insertions: 699 Deletions: 930 Substitutions: 2759
18
19
To use this model with pocketsphinx, just provide it's location with
20
-hmm pocketsphinx option.For information on how to use this model with
21
sphinx4 see the documentation
22
23
http://cmusphinx.sourceforge.net/sphinx4/doc/UsingSphinxTrainModels.html
24
25
Remember that this model uses MLLT and it requires both frontend
26
modification. You need to use transform with LDA component and you need
27
to adjust melfilter parameters.