Phoneme classification is the process of finding the phonetic identity of a short section of a spoken signal. Performances of existing classification techniques are often insufficient, since they rely on Euclidean distances between spectral and temporal features, whereas the relevant features lie in a non-linear manifold. In this work, we propose to integrate into the phoneme classification a non-linear manifold learning technique, namely "Diffusion maps". Diffusion maps builds a graph from the feature vectors and maps the connections in the graph to Euclidean distances, so using Euclidean distances for classification after the non-linear mapping is optimal. We show that Diffusion maps allows dimensionality reduction and improves the classification results.