In this paper, a method is described to acquire speech units automatically which have almost the same duration length as a syllable (or a phoneme). Optimal speech units are obtained following the procedure that the number of units is given beforehand and the pattern of units is updated successively based on a minimum distortion criterion. To investigate whether the speech units used in this procedure correspond to syllables, speech units were acquired automatically on condition that both the number and length of units are equivalent to syllables. As a result, acquired speech units corresponded to about half of syllables. When these units were used for 216 word recognition experiments, the obtained recognition rate was higher than that for the syllable-based method. And a word recognition rate of about 98% or so was achieved by increasing the number of units. At this time, the bit rate was still less than 60 bps, so this method may be applied to the model for ultra very low bit coding.