On-line Emotion Recognition in a 3-D Activation-Valence-Time Continuum using Acoustic and Linguistic Cues

For many applications of emotion recognition, such as virtual agents, the system must select responses while the user is speaking. This requires reliable on-line recognition of the user’s affect. However most emotion recognition systems are based on turnwise processing. We present a novel approach to on-line emotion recognition from speech using Long Short-Term Memory Recurrent Neural Networks. Emotion is recognised frame-wise in a two-dimensional valence-activation continuum. In contrast to current state-of-the-art approaches, recognition is performed on low-level signal frames, similar to those used for speech recognition. No statistical functionals are applied to low-level feature contours. Framing at a higher level is therefore unnecessary and regression outputs can be produced in real-time for every low-level input frame. We also investigate the benefits of including linguistic features on the signal frame level obtained by a keyword spotter.

Perception of emotion: towards a realistic understanding of the task

A decade ago, perceiving emotion was generally equated with taking a sample (a still photograph or a few seconds of speech) that unquestionably signified an archetypal emotional state, and attaching the appropriate label. Computational research has shifted that paradigm in multiple ways. Concern with realism is key. Emotion generally colours ongoing action and interaction: describing that colouring is a different problem from categorizing brief episodes of relatively pure emotion. Multiple challenges flow from that. Describing emotional colouring is a challenge in itself. One approach is to use everyday categories describing states that are partly emotional and partly cognitive. Another approach is to use dimensions. Both approaches need ways to deal with gradual changes over time and mixed emotions. Attaching target descriptions to a sample poses problems of both procedure and validation. Cues are likely to be distributed both in time and across modalities, and key decisions may depend heavily on context. The usefulness of acted data is limited because it tends not to reproduce these features. By engaging with these challenging issues, research is not only achieving impressive results, but also offering a much deeper understanding of the problem.