Next generation sequencing technologies give access to large sets of data which are extremely useful in the study of microbial diversity based on 16S rRNA gene. However, the production of such large datasets is not only marred by technical biases and sequencing noise but also increases computation time and disk space use. To improve the accuracy of OTUs predictions and overcome both computations, storage and noise issues, recent studies and tools suggested removing all single reads and low abundant OTUs, considering them as noise. Although the effect of applying an OTU abundance threshold on α and β diversity has been well documented, the consequences of removing single reads have been poorly studied. Here we test the effect of singleton read filtering (SRF) on microbial community composition using in-silico simulated datasets as well as sequencing data from synthetic and real communities displaying different levels of diversity and abundance profiles. Scalability to large datasets is also assessed using a complete MiSeq run. We show that SRF drastically reduces the chimera content and computational time, enabling the analysis of a complete MiSeq run in just a few minutes. Moreover, SRF accurately determines the actual community diversity: the differences in α and β community diversity obtained with SRF and standard procedures are much smaller than the intrinsic variability of technical and biological replicates.