A nasopharyngeal swab is a sample used for the diagnosis of SARS-CoV-2 infection. Saliva is a sample easier to obtain and the risk of contagion for the professional is lower. This study aimed to evaluate the utility of saliva for the diagnosis of SARS-CoV-2 infection. This prospective study involved 674 patients with suspected SARS-CoV-2 infection. Paired nasopharyngeal and saliva samples were processed by RT-qPCR. Sensitivity, specificity, and kappa coefficient were used to evaluate the results from both samples. We considered the influence of age, symptoms, chronic conditions, and sample processing with lysis buffer. Of the 674 patients, 636 (94.4%) had valid results from both samples. The virus detection in saliva compared to a nasopharyngeal sample (gold standard) was 51.9% (95% CI: 46.3%–57.4%) and increased to 91.6% (95% CI: 86.7%–96.5%) when the cycle threshold (Ct) was ≤ 30. The specificity of the saliva sample was 99.1% (95% CI: 97.0%–99.8%). The concordance between samples was 75% (κ = 0.50; 95% CI: 0.45–0.56). The Ct values were significantly higher in saliva. In conclusion, saliva sample utility is limited for clinical diagnosis, but could be a useful alternative for the detection of SARS-CoV-2 in massive screening studies, when the availability of trained professionals for sampling or personal protection equipment is limited.