This paper describes methods to improve the performance of English phoneme recognition from a linguistic view points. The methods include exploiting time duration information in hidden Markov model (HMM), intrinsic feature space for vowel and consonant recognition respectively, and finally integration of the vowel/consonant recognition results for detail phoneme recognition. As a first step toward the application of these methods, a real time continuous speech recognition system is developed whose software is running on transputer boards plugged into IBM PC.