David Huggins-Daines (PocketSphinx developer) has an article in his wiki comparing Sphinx 3.0 and HTK speech recognition engines.  From the article:

The short story is that, after controlling for the beam widths, language model weights, number of tied states and Gaussians, and the acoustic features, Sphinx 3.0 is actually a bit more accurate than [HTK's] HVite. This is not a very meaningful test, because both are extremely slow (over 5xRT [Real Time] on a 2.8GHz Pentium4). However, it does show that there is nothing inherently wrong with the Sphinx trainer, at least on a basic maximum-likelihood training task.

The Julius speech recognition engine uses HTK acoustic models.

--- (Edited on 7/22/2008 2:43 pm [GMT-0400] by kmaclean) ---