Unveiling Hidden Factors: Explainable AI for Feature Boosting in Speech Emotion Recognition

Explainability improves both performance and trust: The study demonstrates that using Shapley values to understand which features matter most in emotion recognition systems creates a "balance between model performance and transparency," suggesting that making AI systems more interpretable doesn't require sacrificing accuracy.

Removing unnecessary information makes systems work better: By identifying and eliminating irrelevant and redundant features from high-dimensional datasets, the researchers show that simpler, more focused models can outperform complex ones—a finding relevant to responsible AI's goal of building efficient systems that are easier to audit and maintain.

Understanding "why" enables real-world deployment: The paper emphasizes that comprehensive understanding of model predictions is crucial for applications in mental health, education, and human-computer interaction, where stakeholders need to know not just what emotion was detected, but why the system made that determination.

Iterative refinement creates accountability: The proposed feature evaluation loop using Shapley values represents a systematic, repeatable process for improving models, which supports responsible AI practices by enabling ongoing monitoring and adjustment rather than one-time deployment.

Validated across multiple datasets: Testing on four different emotional speech databases (TESS, EMO-DB, RAVDESS, and Surrey) suggests the approach is robust and generalizable, reducing the risk of systems that work only in narrow conditions.