I think the engine takes quite some time to load, it would have to be preloaded some time before that or loaded on first press, then just called (sort of a mute/unmute). The easiest solution I think for the contact names: a dynamically built (the user would choose the names he wants to voice call) jsgf grammar since that would decrease the errors dramatically and hopefully remove the training processes which is abit complex in sphinx as far as i tried.

I know a bit about this as I was working on an AI/Voice Assistant (both on PC using Java, and on Maemo). No solid code yet, and I can't find the time to work on it. However, I'll try to help out as much as I can.

Cheers for the initiative OP