Improving Cross-Corpus Speech Emotion Recognition with Adversarial Discriminative Domain Generalization (ADDoG)

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• Automatic speech emotion recognition (SER) is crucial for user understanding but struggles with cross-dataset generalization.
• Existing adversarial methods for generalized SER representations often face convergence issues and are limited to laboratory data.

Purpose of the Study:

• To introduce novel methods, ADDoG and MADDoG, for robust cross-dataset generalization in speech emotion recognition.
• To address convergence issues and expand applicability to diverse, real-world datasets.

Main Methods:

• Developed Adversarial Discriminative Domain Generalization (ADDoG) using a 'meet in the middle' approach to align dataset representations.
• Introduced Multiclass ADDoG (MADDoG) to extend the method to multiple datasets simultaneously.
• Evaluated methods on cross-corpus speech emotion recognition tasks.

Main Results:

• ADDoG and MADDoG demonstrated consistent convergence.
• Achieved significantly improved cross-dataset generalization, especially without target dataset labels.
• Showed performance improvements over state-of-the-art methods, including with in-the-wild data.

Conclusions:

• ADDoG and MADDoG offer effective solutions for generalized speech emotion recognition.
• The methods show promise for reducing unwanted variation in speech data and can be applied to other domains.