Hey!
You haven't mentioned which approach does your speech recognizer take; if
its DTW, NN or HMM.
If its HMM, you can model the noisy portions as a state (in your word lattice if
you have one, or just fit stick it in appropriately). I'm sure you can
incorporate the same in the other approaches as well.
However, silence/noisy model estimation is unreliable (as the noise environment
might keep changing : babble noise, cocktail party env.) so you would want to
use a noise canceller as a preprocessor.
Again, if you are working with a clean speech environment, forget about all
these and just introduce a silence model :)
Regards,
Abhilash.