The filter bank methods have been a popular non-parametric way of computing the complex amplitude spectrum. So far, the length of the filters in these filter banks has been set to some constant value independently of the data. In this paper, we take the first step towards considering the filter length as an unknown parameter. Specifically, we derive a very simple and approximate way of determining the optimal filter length in a data-adaptive way. Based on this analysis, we also derive a model averaged version of the forward and the forward-backward amplitude spectral Capon estimators. Through simulations, we show that these estimators significantly improve the estimation accuracy compared to the traditional Capon estimators.