Building compact language models for medical speech recognition in mobile devices with limited amount of memory