Language models are derived from large corpora of text or by design. It may be necessary or advisable to adapt these models somewhat before using them with the recognizer.

The language model extensions provide for a number of rather simple mechanisms to adapt an original language model, including:

explicit probabilistic modeling of silence model in between words
explicit probabilistic modeling of filler/garbage model(s) in between words
setting properties for a novel word in relation to the UNK (unknown class)
setting properties for a novel word in relation to another known word

The language model extensions are grouped into a LM-extension file.

For more details, see cwr_lm_ext.c

Limitations

For the time being Language Model Extensions is only supported for Statistical Language Models