To evaluate the effects of record length selection on the accuracy of spectral estimates of heart rate variability (HRV), a simulation study was carried out using a set of 58 signals obtained by autoregressive (AR) fitting a representative sample of real HRV signals. Four record lengths of 180, 300, 420, and 540 s were considered. Spectral estimation was performed by both the Blackman-Tukey (B-T) and AR methods. Accuracy was assessed for: 1) point spectral estimates, by computing the normalized averaged bias (NAB) and variance (NAV); and 2) the most commonly used spectral parameters [total power (TP) and the powers in the bands: very low frequency (VLF) (0 divided by 0.04 Hz), low frequency (LF) (0.04 divided by 0.15 Hz), and high frequency (HF) (0.15 divided by 0.45 Hz)], by computing the normalized bias (NB) and variance (NV). The results are: whatever the record length considered, the 90th percentiles (90P) of the NAB were < 10%, whereas those of the NB were < 9% for TP, LF, and HF powers, and < 14% for the VLF power, in both methods. The NAV was proportional to the reciprocal of record length, showing high 90P values for the shortest record length (26.4% for B-T and 44.2% for AR). The NV showed the same trend but 90P values were much lower (< 8% for TP, LF, and HF powers and < 19% for VLF power, in both methods). In the final part of the paper a procedure for the computation of approximate upper bounds of the relative absolute error of spectral measures at each record length, based on the knowledge of the NB and NV, is presented.