I'm working on this. Pocketsphinx's default model (hub4wsj) was generated from a large number of sentences from the Wall Street Journal. The problem here is that while this gives a good foundation of grammar, it has many words unlikely to be used in normal speech (such as 'Uzbekistan'). These other words make for poor accuracy. On the other hand, one can hand-build a list of words and a simple grammar that works well for voice commands, but not well for more general speech (as we want Saera to handle). I think the best option may be to make a new large dictionary, but create it from a more relevant data source, such as TMO for example. In the meantime I'm just using the default wsj model until I can replace it with something better.