Machine‐Based Modelling
Björn Wolfgang Schuller, Anton M. Batliner · 2013
This chapter deals with the actual machine-based modelling once a feature representation has been found. It starts with the feature relevance analysis, leading on to the actual machine learning. The chapter confines to the approaches encountered most often in the field of computational paralinguistics today. Beyond the machine learning algorithms presented and the variations thereof, there are almost infinitely many others. However, those chosen also present a good balance of basic methods, including static modelling of single feature vectors such as after extraction of supra-segmental features, as well as dynamic approaches that are suitable for modelling on a frame-by-frame level. This choice is motivated by the fact that they are the quasi-standard in many speech processing tasks. The chapter includes classification and regression to allow for discrete class-based and continuous modelling. Finally, it deals with testing protocols, discussing partitioning and balancing of data, performance measures, and also result interpretation.