Sphinx 3.4 coming out soon...

So I just attended a talk by Arthur Chan, formerly of SpeechWorks (now ScanSoft), who is now staff at Carnegie Mellon, and he's working on a new version of the Sphinx 3 recognizer (Sphinx 3.4). It's a very sophisticated approach using a tree-based lexicon (which improves speed, and marginally reduces accuracy of words that are slurred or cut off, since the language model cannot boost confidence scores with this approach), and a continuous HMM (Hidden Markov Model).

Results seem promising -- it is significantly faster than the original Sphinx 3 (which had trouble running in real time except on very very fast machines), for a marginal word-error-rate increase.

We'll need some testers and developers on this soon -- ping me if you'd like to play around with what will be, the best open-sourced (BSD license) recognizer around. I'll be extending a hand to help with this effort some too, though more so after I graduate and have more time to do so.

Kudos to Arthur, Ravi, and Jahanzeb for this effort! And yes, on my agenda in the future (6-12 months from now) will be Ariadne + Sphinx 3.x support. Too many things to do right now like graduating, going to Germany, then off to the Bay Area again (assuming I get offers from at least one of the two jobs I'm seriously looking at).