4.2.7 Network Decoding:
Configuration File
The configuration file consists of parameters more technical than those
found in the parameter file discussed in the previous section. The
parameters in this file affect certain aspects of how the language model is
used by the recognizer. This
section will explain each line of the configuration file used for the
experiment in
Section 4.2.5.
Open the configuration file
config.sof.
@ Sof v1.0 @
@ SearchLevel 0 @
Like the parameter file, the first two lines of the configuration file
allow the recognizer to interpret the following lines as configuration
parameters. These lines must be included in order for the recognizer to
accept the configuration file. The second line, @ SearchLevel 0 @,
means that the following parameters apply to level 0 of the language model,
or in most cases, the state level.
# beam pruning parameters
#
Beam pruning is an efficient technique that removes low scoring hypotheses
early in the search process so the computational resources associated
with those hypotheses can be used for more promising paths.
This line tells the recognizer to use beam pruning for the current
experiment.
Beam-width is the width (defined by the threshold) relative to the
maximum path score at any instant of time. Paths within the search
that fall within this beam-width are kept. The rest are pruned.
# word insertion penalty
#
symbol_penalty = 0;
This penalty is added to the path score everytime a symbol is
hypothesized. Increasing this value will cause the the recognizer to
be more strict about which paths it keeps and which ones are pruned.
Transitions are the probabilities for the HMM's. Think of these as the
probabilities of going from one state to another in a finite state
machine.
This is a scaling factor for transitions. A scaling factor of 0 means
that the probabilities will be disregarded. For TIDigits experiments,
these probabilities usually don't influence the experiment.
This line indicates the end of the parameters for the state level, and
the beginning of the parameters for level 1, or in this case, the
phone level.
# context dependency parameters
#
For this experiment, we're using context-dependent phones. In this case
the recognizer requires additional information about the type of phones
we are using.
use_symbol_context = true;
left_context_length = 1;
right_context_length = 1;
The first parameter simply indicates that we are using context-dependent
phones. The left and right context length parameters are set to 1 which
means we are using phones with both a left and a right context. In other
words, this experiment uses triphones.
Once again, it's very important that the parameters in this file agree
with the language model used for the particular experiment.
|