Some speech recognition companies have trained Acoustic Models using HTK and run them under SPHINX (before HDecode was available). The means/variances/transition probability conversions should be relatively straightforward to accomplish. Requirements:
1) develop some GPL code to do the acoustic model conversion (lets assume ARPA format LMs)
From wout's home page:
HTK to Sphinx 3 model converter
In june 2007, I tried to compare the speed and accuracy of HTK and Sphinx 3 (and ISIP Prototype and Julius) with each other. Because I had models for HTK and didn't want to have to create new models for Sphinx 3, I started working on an HTK to Sphinx 3 model converter.
I didn't have enough time to complete the comparison, but I did almost finish the converter. I say "almost", because although the converter can convert my models without problems and the Sphinx 3 accepts the models, I didn't have time to actually use the models to decode audio. This means that I don't know if there are any non-obvious bugs left in the conversion process and I don't know whether Sphinx 3 can achieve the same performance with converted HTK models as HTK can (i.e. whether the conversion process creates a loss in accuracy).
I won't be developing the converter anymore, but anyone who is interested may download, use and modify the converter for their uses. The converter can be found here.