A Survey on the Semi Supervised Learning Paradigm in the Context of Speech Emotion Recognition

The area of Automatic Speech Emotion Recognition has been a hot topic for researchers for quite some time now. The recent breakthroughs on technology in the field of Machine Learning open up doors for multiple approaches of many kinds. However, some concerns have been persistent throughout the years where we highlight the design and collection of data. Proper annotation of data can be quite expensive and sometimes not even viable, as specialists are often needed for such a complex task as emotion recognition. The evolution of the semi supervised learning paradigm tries to drag down the high dependency on labelled data, potentially facilitating the design of a proper pipeline of tasks, single or multi modal, towards the final objective of the recognition of the human emotional mental state. In this paper, a review of the current single modal (audio) Semi Supervised Learning state of art is explored as a possible solution to the bottlenecking issues mentioned, as a way of helping and guiding future researchers when getting to the planning phase of such task, where many positive aspects from each piece of work can be drawn and combined.