The extraction of a clean fetal electrocardiogram (fECG) from non-invasive recordings is still an open research issue. Despite the number of techniques developed in the literature for fECG extraction, the signal is still affected by a low signal-to-noise ratio (SNR). In this work, different wavelet-based post-processing approaches for the enhancement of the fECG were evaluated. A dataset composed of twenty signals from ten healthy pregnant women, recorded between the 21st and the 27th week of gestation at the San Michele Hospital in Cagliari (Italy), was adopted. fECG extraction was performed by using a multireference QR-decomposition-based recursive least squares adaptive filter as prototypical separation algorithm. Then, all signals were decomposed with the stationary wavelet transform (SWT) and wavelet packet transform (SWPT). In both cases, a 7-level decomposition with Haar mother wavelet and hard-thresholding was selected. Two different thresholds from the literature were tested: the first one is level-independent (Minimax) while the other one is level-dependent. The latter was adapted to be exploited on SWPT. This investigation was meant to perform a targeted denoising based on the fECG spectral bands of major interest, with coarser (SWT) or finer (SWPT) granularity. In this work, the enhancement of the fetal QRS complex was systemati-cally analyzed by computing the SNR before and after wavelet post-processing. Moreover, the accuracy of a fetal QRS peak detector (JQRS, from Physio Toolkit) was evaluated on such signals as a measure of the post-processing performance. The comparative analysis revealed how the SWT outperforms the more complex SWPT, regardless the thresholding approach. The SWT always im-proved the SNR value, reaching a mean increase of 7.5 dB (vs. 4.13 dB of SWPT), and enhanced the fetal QRS complexes with respect to the noise, leading to a better accuracy in the QRS detection (52 % vs. 44 %).