This paper describes how to represent varations of feature parameters to improve recognition of continuous speech. For speech recognition, feature parameters, which are formant frequencies, pitches, logarithmic energies and zero crossing retes are used in general. But, their values and variations depend on speakers, for example disparities between man and woman, and on their age. It is difficult to decide a priority the value of the variation width. Hence, we try to represent this variation by introducing fuzziness and recognize a continuous speech by fuzzy inference using fuzzy production rules.