Expression recognition from non-frontal faces is a challenging research area with growing interest. In this paper, we explore discriminative learning of Gaussian Mixture Models for multi-view facial expression recognition. Adopting the BoW model from image categorization, our image descriptors are computed using Soft Vector Quantization based on the Gaussian Mixture Model. We do extensive experiments on recognizing six universal facial expressions from face images with a range of seven pan angles (−45° ∼ +45°) and five tilt angles (−30° ∼ +30°) generated from the BU-3dFE facial expression database. Our results show that our approach not only significantly improves the resulting classification rate over unsupervised training but also outperforms the published state-of-the-art results, when combined with Spatial Pyramid Matching.