Adapting Multiple Distributions for Bridging Emotions from Different Speech Corpora.

Area of Science:

• Speech Emotion Recognition (SER)
• Machine Learning
• Signal Processing

Background:

• Cross-corpus SER faces performance degradation due to feature distribution mismatches between source and target speech data.
• Existing SER methods struggle when applied to datasets from different speech emotion corpora.

Purpose of the Study:

• To propose a novel transfer subspace learning method, Multiple Distribution-Adapted Regression (MDAR), to address the challenges of cross-corpus SER.
• To bridge the feature distribution gap between different speech emotion corpora.

Main Methods:

• MDAR learns a projection matrix to relate source speech features to emotion labels.
• A novel regularization term, Multiple Distribution Adaptation (MDA), with marginal and conditional operations, ensures applicability to target speech samples.
• The method enables emotion label prediction for target samples using only source label information.

Main Results:

• MDAR demonstrated superior performance on extensive cross-corpus SER tasks using EmoDB, eNTERFACE, and CASIA corpora.
• The proposed method outperformed state-of-the-art transfer subspace learning techniques.
• MDAR also showed better results than several deep transfer learning methods in cross-corpus SER.

Conclusions:

• MDAR effectively overcomes feature distribution mismatches in cross-corpus SER.
• The proposed method offers a robust solution for SER tasks involving diverse speech emotion datasets.
• MDAR represents a significant advancement in transfer learning for speech emotion recognition.