An efficient clustering algorithm called ECA based on long sequences of speech data is proposed for designing vector quantizer. The ECA algorithm, which uses a new splitting scheme and applies the K-means method only in a subset of the training data, greatly reduces the computation requirement. Its special use in phoneme speech synthesis is investigated. Comparisons between the new algorithm ECA and some algorithms presented earlier like LBG, MKM are made based on 12,000 frames of speech data. Simulation results show the advantages of ECA when both the performance and the computation time are considered.