Data-based control of the search space generated by multiple knowledge bases for speech recognition.