Most FFT machines today compute an estimate of the frequency response function, H(f), by the cross-spectral density of input to output divided by the power spectral density of the input. This estimator is contaminated by noise at the input. One uses the coherence function to help measure the level of contamination. However, the coherence function detects, among other things, noise at both the input and output. Several alternate methods are proposed for the computation of the frequency response function. One generates more accurate estimates at resonance, one has half or less of the contamination contained in the present methods, and the last one proposes to eliminate the biasing contamination all together.