The author presents a self-learning neural tree network (SL-NTN) for classification of speech features into phones. The SL-NTN employs a farthest-neighbor fuzzy-clustering algorithm to establish the intra-class correlation among speech phones, thus splitting the phone in such a way as to maximize the recognition performance while reducing the computational complexity. When evaluated on the 61 phones of the TIMIT database, the SL-NTN has been shown to provide an optimal trade-off between computational complexity and recognition performance. It also provides insight into the relationship among the applied speech patterns.<<ETX>>