Progressively Discriminative Transfer Network for Cross-Corpus Speech Emotion Recognition.

Area of Science:

• Speech Emotion Recognition (SER)
• Machine Learning
• Domain Adaptation

Background:

• Cross-corpus SER faces performance degradation due to domain shift between training and testing data.
• Existing domain adaptation methods may impair emotion discrimination by excessively narrowing feature distributions.
• A gap exists in methods that simultaneously address domain mismatch and preserve emotional distinctiveness in SER.

Purpose of the Study:

• To propose a novel network, the Progressively Discriminative Transfer Network (PDTN), for robust cross-corpus SER.
• To enhance emotion discrimination capabilities of speech features while mitigating domain shift.
• To improve the performance of SER models on unseen target domains.

Main Methods:

• Introduced the Progressively Discriminative Transfer Network (PDTN) architecture.
• Designed emotion discriminant loss (Ld) incorporating valence-aware (Lv) and emotion-aware (Lc) center losses.
• Implemented multi-layer distribution alignment loss (La) and combined it with cross-entropy loss (Le) for optimization.

Main Results:

• The PDTN effectively enhances emotion discrimination while reducing domain discrepancy.
• Experimental results on six cross-corpus tasks across Emo-DB, eNTERFACE, and CASIA datasets demonstrate superior performance.
• PDTN outperforms existing state-of-the-art methods in cross-corpus SER.

Conclusions:

• The proposed PDTN successfully addresses the challenges of cross-corpus speech emotion recognition.
• The novel loss functions (Ld and La) are crucial for balancing domain alignment and emotion discrimination.
• PDTN offers a promising approach for developing more generalizable and accurate SER systems.