Classifier Subset Selection for the Stacked Generalization Method Applied to Emotion Recognition in Speech.

Area of Science:

• Computer Science
• Artificial Intelligence
• Speech Processing

Background:

• Speech emotion recognition (SER) is crucial for human-computer interaction.
• Existing multi-classifier systems like stacking generalization can be improved for SER tasks.

Purpose of the Study:

• To introduce a novel supervised classification paradigm, classifier subset selection for stacked generalization (CSS stacking).
• To enhance stacking generalization by integrating an estimation of distribution algorithm (EDA) for optimal base classifier subset selection in SER.

Main Methods:

• Developed CSS stacking by integrating an EDA into the first layer of a bi-level multi-classifier system.
• Evaluated CSS stacking on the RekEmozio and Berlin Emotional Speech datasets using spectral, quality, prosodic, and acoustic features (eGeMAPS).
• Compared CSS stacking performance against standard classifiers and traditional stacking generalization.

Main Results:

• CSS stacking demonstrated strong performance across various configurations and datasets.
• The proposed method showed significant improvements over single classifiers and standard stacking.
• Performance was validated using different feature sets and classifier combinations.

Conclusions:

• CSS stacking is an effective approach for improving speech emotion recognition.
• The integration of EDA for classifier subset selection enhances the performance of stacking generalization.
• The paradigm shows promise for real-world SER applications.