Surface Enhanced Raman Spectroscopy (SERS) technique is a fast, convenient and nondestructive optical diagnose method that can provide specific fingerprints of biomolecular compositions and structures. Human saliva contains many biomarkers that can reflect health condition. In this paper, we employed SERS methods based on nanometer silver colloid to obtain high quality SERS of 208 saliva samples from 46 patients with nasopharyngeal carcinoma (NPC), 30 patients with rhinitis and 48 healthy volunteers. The results showed there were great differences in SERS pattern among these three subjects. Principal component analysis-linear discriminant analysis (PCA-LDA) was utilized to analyze and classify the saliva spectra from NPC, rhinitis and healthy subjects. The diagnostic sensitivity was 89.5%, 93.0% and 100% respectively, and meanwhile the specificity was 94.6%, 96.7% and 97.7% respectively. The research showed that saliva SERS combined with PCA-LDA can rapidly and effectively distinguish three saliva groups, and this technique has great potential to be a non-destructive and clinical screening method for nasopharyngeal carcinoma detection.