Large-vocabulary continuous-speech recognition using partitioned graph search