Automatic generation of subword units for speech recognition systems