Multi-Path and Group-Loss-Based Network for Speech Emotion Recognition in Multi-Domain Datasets.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• Speech Emotion Recognition (SER) models face challenges with limited labeled datasets and poor generalization to new domains.
• Real-world SER applications require models that can adapt to diverse, unseen data.
• Existing SER models often struggle with cross-domain performance.

Purpose of the Study:

• To propose a novel Multi-Path and Group-Loss-based Network (MPGLN) for robust speech emotion recognition.
• To enhance the multi-domain adaptation and generalization capabilities of SER models.
• To address the limitations of current SER approaches in handling diverse emotional speech data.

Main Methods:

• Developed a Multi-Path and Group-Loss-based Network (MPGLN) incorporating a bidirectional LSTM temporal feature generator and VGGish-based feature extractor.
• Implemented simultaneous learning using multiple losses based on discrete and dimensional emotion labels.
• Evaluated the model on multi-cultural datasets, including the Korean Emotional Speech Database (KESD) and the Interactive Emotional Dyadic Motion Capture (IEMOCAP) database.

Main Results:

• The MPGLN demonstrated significant improvements in multi-domain adaptation and domain generalization.
• Achieved a 3.7% improvement in F1 score for multi-domain adaptation compared to a baseline model.
• Showcased a 3.5% improvement in F1 score for domain generalization.

Conclusions:

• The proposed MPGLN effectively supports multi-domain adaptation in speech emotion recognition.
• The MPGLN architecture reinforces model generalization, making SER models more applicable to real-world scenarios.
• This approach offers a promising solution for developing more versatile and accurate SER systems.