You do not have to use any noice reduction algorithms.
If you do use one, make sure you apply it to both the training audio and recognized audio. Otherwise you may notice a drop in accuracy.
Depending on the algorithm, you may be losing some important data if you use it. The sound may seem better for you, as a human, but it may become harder to recognize for Sphinx.