Hyperspectral imaging (HSI) helps in nondestructive forensic stain analysis through precise spectral characterization of materials. Still, the problem of distinguishing visually corresponding stains of beverages and subspace spectral variability exists for high-dimensional spectral data. This paper proposes an autoencoder-based deep spectral learning approach for the classification of hyperspectral datasets of beverages. Spectral signatures of individual pixels, which consist of 204 adjacent spectral bands, were obtained from hyperspectral datasets of nine different beverages: Papaya, Coffee, Pomegranate, Orange, Tea, Wine, Whisky, Rum, and Brandy. The research employs a fully linked autoencoder to acquire low dimensional spectral embeddings from high-dimensional spectral data, thereby minimizing redundancy within the dataset. These embeddings were further used for classification purposes using a one-dimensional convolutional neural network (1D-CNN). Performance evaluation of the proposed approach showed significant enhancements in class separability and test classification accuracies of 96.9 %, when compared to direct deep learning on spectral characteristics. The results confirm the feasibility of learning low-dimensional spectral embeddings for non-destructive forensic analysis of beverages.