Exploring latent structure in expressive speech

A large part of research on expressive speech focuses on a fixed palette of “fully-blown” emotions leaving a large set of interesting applications unaddressed. The work described here adopts a more generic approach to expressive speech analysis. By recognizing that affect is manifested in speech through a rich, diverse set of patterns involving various surface features, it seeks to reveal underlying structure and latent components in these patterns which could lead to a fuller understanding of the expressive techniques employed to convey emotion. This places emphasis not on the emotion itself but on expressive speaking style. To this end, it exploits the underlying structure in the expressive speech data to derive a set of latent emotional features. This not only helped in removing a large part of the redundancy in the measured acoustic features, but also revealed latent components in speech that could more efficiently describe affective content. This approach seems more appropriate for applications where expressivity and affect cannot be sufficiently described in terms of a (closed) set of primitive (and archetypical) emotions, as is the case of child-directed speech, narratives and storytelling applications.