Subsymbolic Approach to Word Modeling for Domain Specific Speech Recognition