| |
-
Continued from Flat-start:
Once we have trained decent seed models in the
flat-start
stage of the training process, we can then train the models on
a larger training set.
For this we simply need to replace the
feature file list
and the
model label file
with the appropriate files in the parameter file and use:
hmm_train
-p
params.text
-c CI
-
Creating the Short Pause Model:
In the
data preparation
section of this tutorial, we saw that the short pause (sp)
model was a dummy model. In this stage we train this model.
Typically, the "sp" model has a single state and is tied to the
middle state of the "sil" model. This can be simply done by setting
the central state of the sp model in the model definition file to
the central state of the silence model from the last step.
The steps involved for our example follow:
Assign a new transition matrix (index 30 is the next one in
this case) for the "sp" model.
In the
transitions file
generated at the end of the last training iteration increment
the transition matrix count by one and add a new 3x3 matrix
with index chose above (index 30). Initialize it to:
0.000000e+00 5.000000e-01 5.000000e-01
0.000000e+00 5.000000e-01 5.000000e-01
0.000000e+00 0.000000e+00 0.000000e+00
-
Training with the Short Pause Model:
Once the "sp" model has been initialized, we use
word labels
of the training data to train the models. Note that while
training with word transcriptions, a couple of changes need to
be made to the
parameter file
to be used with
hmm_train.
-
Change the mlf_file tag to point to the
word label file.
-
Set the mlf_mode tag to "word_mlf" instead of
"model_mlf".
With these changes made, the new parameter file should appear
similar to
this.
Use
hmm_train
to start training. About four iterations of training should be
done in this stage. The process is shown as a flow graph below.
prev
next
top
|
|
|
|
