Cross-Corpus Speech Emotion Recognition Based on Transfer Learning and Multi-Loss Dynamic Adjustment.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• Cross-corpus speech emotion recognition (SER) faces challenges due to mismatched feature distributions between training and testing datasets.
• Classical algorithms often struggle to achieve optimal performance in cross-corpus SER scenarios.

Purpose of the Study:

• To propose an effective algorithm for cross-corpus SER that addresses feature distribution mismatches.
• To improve the generalization ability and recognition accuracy of SER systems when applied to unseen speech corpora.

Main Methods:

• A novel deep network model combining a deep auto-encoder and fully connected layers was developed to enhance feature representation.
• Global and subdomain adaptive algorithms were employed for effective feature transfer across different speech corpora.
• Dynamic weighting factors were introduced to balance multiple loss functions, preventing optimization offset during model training.

Main Results:

• The proposed transfer learning and multi-loss dynamic adjustment (TLMLDA) algorithm demonstrated excellent recognition results on the Berlin, eNTERFACE, and CASIA speech corpora.
• The algorithm achieved competitive performance compared to most state-of-the-art methods in cross-corpus SER tasks.
• The TLMLDA algorithm effectively improved the generalization ability of the SER system.

Conclusions:

• The TLMLDA algorithm offers a robust solution for cross-corpus speech emotion recognition by effectively handling feature distribution discrepancies.
• The proposed method shows significant potential for real-world applications requiring accurate emotion detection from diverse speech data.
• This research contributes to advancing the field of SER by providing a more generalizable and accurate approach.