Time-varying autoregressive modeling may consider the driving noise variance as a constant. In this work, the properties of the autoregressive driving noise variance of heart rate variability, with different stationary physiological conditions (resting in supine and sitting; exercise) are obtained. The effect of constant variance consideration for ramp exercise and recovery (a nonstationary condition) is also evaluated by the comparison of the time-varying absolute spectral parameters obtained by parametric estimation, allowing or not the modeling of time-varying noise variance, and a non-parametric time-frequency analysis. The driving noise variance presented a direct non-linear relationship with the heart period for the stationary maneuvers (r=0.91), while for the nonstationary condition, the use of a constant driving noise resulted in bias for the estimation of heart rate variability spectral parameters. A time-varying driving noise variance should be considered.