The wavelet transform is a powerful tool for capturing the joint time-frequency characteristics of a signal. However, the resulting wavelet coefficients are typically high-dimensional, since at each time sample the wavelet transform is evaluated at a number of distinct scales. Unfortunately, modelling these coefficients can be problematic because of the large number of parameters needed to capture the dependencies between different scales. In this paper we investigate the use of algorithms from the field of dimensionality reduction to extract informative and compact descriptions of shape from wavelet coefficients. These low-dimensional shape descriptors lead to models that are governed by only a small number of parameters and can be learnt successfully from limited amounts of data. The validity of our approach is demonstrated on the task of automatically segmenting an electrocardiogram signal into its constituent waveform features.