The two decades referred to in the subtitle essentially span the time since the publication of Picard’s foundational Affective computing [1], which began the study of emotion recognition by computers. This paper can therefore be viewed as a comprehensive review of emotion recognition in speech.

The author begins by laying out an overall view of the process. In gross terms, the process has four components. First, one chooses the model for emotions, either discrete classes or a value continuous dimensional view composed of axes for arousal and positivity. Then one acquires labeled data. Following this, features are selected that are then fed into a learning system. Initially, the labeling of the data required extensive human intervention with the ambiguities that this implies, but now systems exist where the machine can learn to label the data with some human intervention.

This is an iterative process where human advice is used to learn labels. Features can be chunks of audio rather than just words. It is also important to take into account the speaker’s states and traits beyond the emotion of interest. The author summarizes the results of recent speech emotion recognition (SER) challenge events in a useful table. Finally, the author considers challenges that the SER community could undertake.

Going beyond the recognition of irony or sarcasm, the author suggests what he calls a “moonshot challenge” to target the actual emotion of the speaker. The review illuminates a fascinating area and leaves the reader eager for more. There is a comprehensive bibliography.