Analysis of surface electromyography (sEMG) signals is a common practice in biomedical applications for recognizing muscle movement, wavelet coefficients obtained from wavelet transform (WT) or wavelet packet transform (WPT) are used as features of the sEMG signal and classified by means of machine learning models. To the best of our knowledge, no study has fully exploited the resemblance wavelet coefficients have to the signal from which they were obtained. In this context, time domain feature extraction on smaller data lengths can be applied directly to approximation and detail coefficients for different decomposition levels. This can be seen as different frequency band filtered versions of the original signal. The aim of this research is to compare time domain feature extraction of wavelet coefficients obtained from WT and WPT against time domain feature extraction for different frequency bands filtered sEMG signals and determine which approach is most suitable for hand movement recognition. To this end, sEMG signals were decomposed using both the WT (level 6, 'db4') and WPT (level 3, 'db4') methodologies to compare results. The comparison criterion reflects the results of the classification of three machine learning models. Results were obtained by performing supervised multiclass classifications of 18 upper limb movements from 40 subjects, retrieved from the 2nd public database generated for the Ninapro Project. The use of a lower number of coefficients can produce similar performance results as shown when comparing WT vs WPT. In the other hand, time domain feature extraction from filtered sEMG signals using wavelet reconstruction produces slightly better performance on classification results at a higher computational cost.