An in vivo study of human oral cancer diagnosis by using autofluorescence spectroscopy is presented. A Xenon-lamp with a motor-controlled monochromator was adopted as the excitation light source. We chose the excitation wavelength of 330 nm, and the spectral measurement range was from 340 nm to 601 nm. A Y-type fiber bundle was used to guide the excitation light, and collect the autofluorescence of samples. The emitted light was detected by a motor-controlled monochromator and a PMT. After measurement, the measured sites were sectioned and sent for histological examination. In total 15 normal sites, 30 OSF (oral submucosa fibrosis) sites, 26 EH (epithelial hyperkratosis) sites, 13 ED (epithelial dysplasia) sites, and 13 SCC (squamous cell carcinoma) sites were measured. The discriminant algorithm was established by partial-least squares (PLS) method with cross-validation technique. By extracting the first two t-scores of each sample and make scattering plot, we found that the samples of different cancerous stages were in grouped distinct locations, except that samples of ED and EH were mixed together. It means that this algorithm can be used to classify normal, premalignant, and malignant tissues. We conclude that autofluorescence spectroscopy may be useful for in vivo detection of early stage oral cancer.