A novel acoustic modeling algorithm that generates non-uniform unit HMMs to effectively cope with spectral variations in fluent speech is proposed. The algorithm is devised for the automatic iterative generation of long-span units for non-uniform modeling. This generation algorithm is based on an entropy reduction criterion using text data and a maximum likelihood criterion using speech data. The effectiveness of the non-uniform unit model is confirmed by a phrase recognition test using an LR parser. Recognition results show that non-uniform unit HMMs achieve higher performance than conventional phoneme-unit HMMs and suggest the potential capacity of non-uniform unit HMMs.