Fuzzy neural networks (FNN) have several features that make them well suited to a wide range of knowledge engineering applications. These strengths include fast and accurate learning, good generalisation capabilities, excellent explanation facilities in the form of semantically meaningful fuzzy rules, and the ability to accommodate both data and existing expert knowledge about the problem under consideration. The paper presents one particular architecture called FuNN and discusses two alternative ways to optimise its structure, namely a genetic algorithm and a method of learning-with-forgetting. The optimised structure has much less connections and can easily be interpreted in terms of fuzzy rules. Such a structure can be effectively used for on-line adaptation which is demonstrated on a phoneme-based speech recognition problem.