| |
-
Triphone Initialization:
We need to first create the definitions of the triphone models
we intend to train. In many cases the user may want to do some
sort of state-tying. A result of model-tying is that there are
a small number of physical models which map to a large number of
logical models.
The
model file
will contain definitions of the physical models only. The
clustered list file
is used to specify the mapping between physical and logical
models when using the decoder as well as the training utility.
We create the model definitions by using the following command
line, assuming we have a clustered phones list. If model-tying
is not done, then the tri_clist_file tag should point to
a file containing the names of all logical models.
init_triphones
-param
itri_params.text
These
are the triphone model definitions which corresponds to using
this
phones file.
We now have to define the mapping between logical and physical
models using the
create_triphone_map utility.
create_triphone_map
-mono
fs_ci_models.text
-clist
triphones.clist
-context cd
-models
tri_models.text
-output
tri_phones.text
We are now ready to move on to training the triphone models.
-
Triphone Training:
Training a cross-word context-dependent triphone system now is
easy. All we need to do is to change the context_mode
tag in the parameter file to "cross_word" and set the other
supporting data like the model definitions, phone mapping
etc. appropriately.
hmm_train
-p
cd_train_params.text
-c CD
A few iterations of the above process and we are all set to use
the triphone models for recognition.
prev
next
top
|
|
|
|
