A technique is described for extracting spectral/temporal features from speech segments such that more emphasis is given to the center of the segment and less to the end regions. A classification technique, called binary-pair partitioning (BPP), is also described. This method partitions an N-way classification task using N*(N-l)/2 elemental classifiers, each of which discriminates one pair of categories. These features and this classification technique resulted in 72.4% accuracy for classification of 16 vowels extracted from the DARPA/TIMIT data base in speaker-independent experiments.