Use of high-level linguistic constraints for constructing feature-based phonological model in speech recognition