In many classication approaches, time series comparison requires the time series to be aligned. Numerous strategies for temporal alignments have been proposed in the literature; they intend to t observations to make compared time series as close as possible (e.g., Rodrigue et al. (2004), Shou et al. (2005), Navarro (2001), Nanopoulos et al. (2001)). In the framework of time series discrimination, this work focuses on learning time series alignments by connecting the commonly shared temporal features within clusters (i.e., higher clusters cohesion), and the greatest dierences between clusters (i.e., higher clusters isolation). A new time series alignments approach supervised by a variance/covariance criterion is proposed. The core of the alignment strategy is based on strengthening or weakening links according to their contribution to the variability within and between classes. To this end, the classical variance/covariance expression is extended to a set of time series, as well as to a partition of time series. Discriminative distances based on the learned alignments are then induced for time series classication. We show, through the carried out experiments, that the learned distances outperform the standard ones for time series classication.