Analysis of Clustering Algorithms for Speech Emotion Recognition

Abstract

Speech emotion recognition (SER) systems are used to automatically identify emotional content in human speech communication. The development of accurate SER models often requires large volumes of labeled emotional speech data. Annotating emotional speech data is both costly and time-consuming, partly due to the sheer volume of data involved and partly due to the subjective nature of detecting paralinguistic properties in audio signals. This study investigates a clustering-based active learning (AL) approach to mitigate the annotation burden while maintaining high model performance. The research particularly examines the impact of different clustering algorithms and distance metrics on the performance of clustering-based AL in SER. The results indicate that mini-batch k-means and BIRCH outperformed the other clustering algorithms in terms of large-scale SER data.