Accurate prediction of shelf-life is critical for reducing food waste and boosting consumer confidence in the food industry. Conventional methods like gas chromatography-mass spectrometry and microbial analyses are commonly used to monitor changes in food quality and safety and predict shelf-life. These methods, while accurate, are expensive, time-consuming, and impractical for real-time applications. Hence, a more affordable, simple, and real-time analytical method was developed in this study and tested with samples of soymilk, leveraging the volatile compounds (VOCs) released during spoilage. A novel gold-coated surface-enhanced Raman spectroscopy (SERS)-active fiber was inserted into the headspace of soymilk in capped glass vials. The SERS fiber was incubated in the headspace of the samples at 25ºC for 30 minutes. SERS spectra were acquired using a Raman microscope. Samples were monitored for two weeks. The SERS-active fiber could capture VOCs from soymilk, especially dimethyl sulfide, a well-known bacterial spoilage VOC, during spoilage. Using convolutional neural networks (CNN) models, the SERS spectral changes strongly correlated with changes in pH (R=0.89, RMSE=0.30), microbial growth (R=0.91, RMSE=0.69log10CFU/ml), electrical conductivity (R=0.92, RMSE=0.07mS/cm), zeta-potential (R=0.94, RMSE=1.26mV) and remaining shelf-life (R=0.95, RMSE=1.30days) in the training samples. Further, using SERS spectra from the headspace of a separate set of soymilk samples, the CNN models could predict the remaining shelf-life (R=0.80, RMSE=2.54days) and microbial growth (R=0.73, RMSE=1.14log10CFU/ml). Therefore, the SERS fiber combined with CNN can predict soymilk quality and shelf-life in real-time. Further studies are needed to advance the practical application of this method to more food products.