Spectral parameters extracted from the heart rate variability signal are obtained on a beat-to-beat basis by means of autoregressive recursive identification. In this paper a whale forgetting window is introduced, instead of the classical exponential one, in order to reduce the noise influence on the estimated parameters. After proper simulation it was found that the whale forgetting window markedly reduces the noise in the identification, but maintains a good response to abrupt changes in the signal. The algorithm was thus applied to the analysis of the HRV data recorded during different transient situations in physiological and pathological conditions. The spectral parameters were obtained on a beat-to-beat basis and their trends were smoother and more accurate with respect to the traditional exponential window also in presence of noise or artifacts in the time series (sudden and short time changes, ectopic beats, etc.), without losing the signal variations of physiological interest.