Advanced parameters for language model building

The offline grammar compiler contains advanced parameters that are used for language model building.

The advanced SLM parameters listed below are complex parameters that are best understood by speech language specialists. These are advanced parameters that require a deep understanding of statistical language models and cannot be easily explained by themselves.

Specify the following advanced SLM parameters:

-ngramlength <positive integer value>
Specify the maximum length of n-grams to count when processing the sentence input. The default is 3. Usually this value is 2 or 3, rarely 4. This field reflects the size of the training data to generate when building the model. The -ngramlength must be greater than or equal to -ngramorder.
-ngramorder <positive integer value>
Specify the order of the n-gram model, which is the maximum length of n-grams to build into the model from the training data. The default is 3; rarely should this value be 4. This field reflects the size of the model; a higher value typically means higher accuracy. The -ngramlength must be greater than or equal to -ngramorder.
-ngramcutoffs <string>
Specify the cutoffs, which is a string containing thresholds for each n-gram level separated by commas; no spaces are allowed. Cutoffs represent the maximum count that is removed, to reduce the size of the generated model. The number of cutoffs must match the -ngramorder, i.e. if you use 3, then you need to specify 3 values, such as 1,1,1 for (unigram, bigram, trigram).
Note: Specify the threshold parameter to control the size of the language model. If all information in the training data is contained in the model (i.e., no thresholds applied), the model size can be prohibitively large for the targeted application, for example, because of memory or CPU limitations. In such situations, the thresholds can be used selectively to remove the information about infrequent events from the model. The higher the value of each threshold, the smaller the model will be. The values have to be increased with caution, if they become too high, it may happen that there will not be enough training data available and the attempt to build the model will end in an error.
-contextsize < positive integer value or zero>
Specify how many cross word phones should be considered. A value greater than 1 (up to 4) slows the compilation time significantly, but may provide a small increase in recognition accuracy. This field is used to mostly determine compilation time. The default is 1.
-insertioncost <positive float value or zero>
Specify the cost of a word insertion. The value can be set from 0.0 to +inf. The value of 0 corresponds to the highest probability of word insertion; increasing this value will cause fewer but longer words to be considered in the model. The default is 0.0.
-silenceloglikelihood <negative float value or zero>
Specify the Log likelihood of silence. The value can be set from 0 to -inf. The value of 0 corresponds to the highest log-likelihood of silence insertion. The default is -0.6.

Terms of use | Feedback

(c) Copyright IBM Corporation 2004, 2009. All rights reserved.
U.S. Government Users Restricted Rights - Use, duplication or disclosure restricted by GSA ADP Schedule Contract with IBM Corp.