This paper describes an optimal statistical method to recognize phonemes in continuous speech. The novelty of this method is to search the most effective acoustic features in each acoustic level using the criterion of mutual information between acoustic feature vectors and phoneme labels assigned to the speech wave. In the proposed method for phoneme recognition using multiple acoustic features, input speech is first classified based on acoustic similarity, and possible phoneme is selected using variable acoustic features hierarchically. On each level of acoustic features including power and its variational pattern, LPC Mel-cepstrum and its pattern of temporal change are precisely evaluated. Multi-level clustering is suitable to discriminate phonemes by detecting the most reliable features in that context and by using the effective combination of various acoustic characteristics.