| |
6.2.4 Network Builder: Acoustic Models
In addition to creating language models, Network Builder can also be
used to create acoustic model files. The steps to doing this are simple.
To begin with, you need a completed language model file, built as explained
in the previous sections. Go to the directory:
$ISIP_TUTORIAL/sections/s06/s06_02_p04/
There should be a few files in this directory, one of them called
lm_phone_jsgf.sof
This should be identical to the language model we created in
Section 6.2.4
We will create an acoustic model file for this language model.
Open the language model with network builder and select the state level.
Next click the Hierarchy menu
and click Symbols. This will give you a table of all the symbols
in this level of the hierarchy. The panel to the right of this table contains
a list of symbol types, one of which is 'attach statistical'. Click the
first symbol and drag your mouse across the remaining symbols. This should
select all of them. Once they are all highlighted, click 'attach
statistical', and then click OK.
Now, go to the File menu and click Save All. A window will pop
up with two text boxes, one specifying where to save the language model
file and the other specifying where to save the statistical model. Save
the statistical model as smp_phone.sof in the current directory. Save
the language model as lm_phone_jsgf.sof.
Click OK and exit Network Builder.
The file Network Builder creates is only a hash table of symbols. We must
now use isip_recognize to complete the acoustic model file by creating
models that correspond to each of the symbols. For this step, we'll use the
"MODEL_CREATION" algorithm of isip_recognize.
There should be a parameter file called
params_mc.sof
in your current directory. The important difference between
this parameter file compared to the ones we've worked with
previously is the algorithm specified as "MODEL_CREATION". This
tells isip_recognize to generate an acoustic model file from
the hash table.
Run the following command:
isip_recognize -param params_mc.sof -verbose ALL
This will produce a new acoustic model file which contains models
corresponding to every symbol in the hash table. This file is
now ready to be used in training. (See
Section 5)
|
| |
|