Building a Gender-Bias-Resistant Super Corpus as a Deep Learning Baseline for Speech Emotion Recognition.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• Speech Emotion Recognition (SER) is crucial for intelligent systems but faces challenges in robustness and bias.
• Current SER standards are outdated, despite advancements in deep learning architectures.
• Existing deep learning models for SER lack thorough examination regarding speaker gender and out-of-distribution data.

Purpose of the Study:

• To create a comprehensive super corpus for SER by merging existing databases.
• To establish a new benchmark for SER using diverse deep learning architectures.
• To investigate and mitigate gender bias and improve generalization in SER models.

Main Methods:

• Construction of a novel super corpus by aggregating data from multiple SER databases.
• Benchmarking the super corpus with various deep learning architectures to set a new performance baseline.
• Implementation of data augmentation strategies to address and reduce inherent data biases.

Main Results:

• Models trained on the super corpus exhibit enhanced generalization and accuracy compared to those trained on individual datasets.
• Training on the super corpus significantly reduces gender bias in SER models.
• Data augmentation proves effective in mitigating biases across genders and emotions, sometimes achieving full debiasing.

Conclusions:

• The developed super corpus provides a robust foundation for advancing SER research.
• Deep learning models trained on the augmented super corpus demonstrate improved fairness and performance.
• Data augmentation is a critical technique for developing unbiased and more accurate SER systems.