This paper investigates the measurement errors involved in measuring frequency response functions from weighted-overlapped segment averaging, a technique that has become a standard in modern spectral analysers due to its computational advantages. A particular attention is paid to leakage errors, for which this procedure has been frequently criticised. Exact and asymptotic expressions for the bias and variance are provided, whose minimisation enables the derivation of the optimal settings to be used with this procedure. Our main finding is that a Half-sine or Diff window with 2/3 overlap achieves the best compromise to reduce leakage errors, and this is independently of the system frequency response function. This conclusion is to be contrasted with the customary habit of using a Harming window with 1/2 overlap. (c) 2007 Elsevier Ltd. All rights reserved.