This paper presents a sparse coding algorithm based on deep auto-encoders for the stylization and the clustering of pitch contours. The main objective of the proposed algorithm is to learn a set of pitch templates that can be easily interpreted by humans and whose combination can approximate efficiently the observed pitch contours. The proposed learning architecture is based on deep auto-encoders, commonly used to learn non-linear and low-dimensional latent representations that approximate the observed data. The proposed deep architecture is based on stacked auto-encoders and the sparsity of the network is investigated in order to learn a more robust and general representation of the pitch contours (dropout, denoising auto-encoder, sparsity regularization). The deep auto-encoding of the pitch contours is illustrated and discussed on the TIMIT American-English speech database † with comparison of other existing stylization and clustering algorithms.