The accuracy of spectral analysis of heart-rate variability performed on annotated RR interval lists obtained from several commercial Holter systems was appraised. Five tape-recorder-based systems (Del Mar 750, Marquette 8000, Oxford Medilog Excel, Remco Cardioline AD 35 and Reynolds Pathfinder PA3) and four solid-state systems (Hewlett Packard 43420B, Marquette Seer, Oxford 6000FD2, Reynolds E-Ram) were considered. Two ECG signals with fixed real morphology but characterized by a different degree of modulation of the RR interval (reduced and normal variability) were fed into the recorders evaluated. The total power and the power in the very low-, low- and high-frequency bands were then estimated on all Holter-generated RR sequences. Spectral analysis was performed by both the autoregressive and fast-Fourier-transform methods. The estimation error of each parameter was statistically characterized and, for tape-recorder-based systems, inferential analysis was used to test for differences between recorders, tapes and times of recording. The centre and dispersion of the estimation error changed markedly from system to system. Some tape-recording systems showed large inter-recorder differences. The degree of spectral distortion was never uniform among selected bands. Solid-state systems performed better than tape-recording ones but both were limited in the accuracy by the quantization of RR interval measurement. The fast Fourier method yielded spectral estimates more stable than the autoregressive method. Our data clearly show that spectral analysis of very low-variability signals may be seriously affected by Holter recording and preprocessing of ECG signals.