Diagnosis in neuro-oncology can be assisted by non-invasive data acquisition techniques such as Magnetic Resonance Spectroscopy (MRS). From the viewpoint of computer-based brain tumour classification, the high dimensionality of MRS poses a difficulty, and the use of dimensionality reduction (DR) techniques is advisable. Despite some important limitations, Principal Component Analysis (PCA) is commonly used for DR in MRS data analysis. Here, we define a novel DR technique, namely Spectral Prototype Extraction, based on a manifold-constrained Hidden Markov Model (HMM). Its formulation within a variational Bayesian framework imbues it with regularization properties that minimize the negative effect of the presence of noise in the data. Its use for MRS pre-processing is illustrated in a difficult brain tumour classification problem.