Automatic learning of lexical representations for sub-word unit based speech recognition systems.