Robust and Reliable Speech Understanding in Restricted Domains