The content of total volatile basic nitrogen (TVB-N) is an important index for evaluation of pork freshness. To achieve rapid and nondestructive detection of pork freshness, we use a light-emitting diode (LED) light source containing optimal characteristic wavelengths to set up a multispectral detection system by selecting optimal characteristic wavelengths related to the content of TVB-N in pork, and the content of TVB-N in pork is detected by the detection system. Firstly, a visible near infrared (VIS-NIR) hyperspectral system is applied to detect pork to acquire hyperspectral reflectance data, and then different preprocessing algorithms including the first derivative (FD), standard normal variable transformation (SNV), and other methods are utilized to build partial least squares regression (PLSR) model of the TVB-N content in pork. Secondly, some variable selection methods including the step wise algorithm (SWA), successive projections algorithm (SPA), and genetic algorithm (GA) are used to screen the characteristic wavelengths related to TVB-N content. PLSR model and multiple linear regression (MLR) model are established by these characteristic wavelengths. The model results are compared to choose optimal characteristic wavelengths. Lastly, the LED light source containing characteristic wavelengths is used in the multispectral detection system to establish PLSR and MLR models, so that the detection of TVB-N content in pork is completed. The results indicate that the screening characteristic wavelengths by using the SWA, SPA, and GA can reflect full spectral information well. The effect of the established model is good, and the number of variable decreases greatly. The LED light source containing characteristic wavelengths can detect the TVB-N content in pork well in the multispectral detection system. The results of the established MLR model are better than those of the PLSR model. The correlation coefficient and square error of calibration (SEC) set of the MLR model are 0.9050 and 3.63×10 -5, respectively, and the correlation coefficient and square error of prediction (SEP) set are 0.9040 and 3.81×10 -5, respectively.