| Resumen |
Speech emotion recognition datasets can have class noise due to subjectivity during the labeling process or because of automatic labeling. Class noise has been neglected until recently in machine learning research. In this work, we study the effects of controlled class noise in 10 datasets from different languages. We find that low levels of class noise (5-10%) do not significantly affect the performance of classifiers, but higher levels of class noise severely impact performance. Support Vector Machines (SVMs) appear to be the best candidate for handling class noise across most datasets and noise levels compared to other traditional algorithms. We propose a preprocessing method that effectively corrects mislabeled samples at moderate and high noise levels, enhancing the model’s performance as measured by balanced accuracy. |
