Strong Generalized Speech Emotion Recognition Based on Effective Data Augmentation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• Limited labeled data hinders Speech Emotion Recognition (SER) model development.
• Data augmentation is crucial for addressing data sparsity in SER.
• Existing research on data augmentation algorithms for SER is scarce.

Purpose of the Study:

• To analyze the effectiveness of classical acoustic data augmentation in SER.
• To propose a generalized SER model leveraging effective data augmentation.
• To enhance model robustness for unseen speakers through adversarial training.

Main Methods:

• A multi-channel feature extractor with sub-networks to capture emotional representations.
• Integration of various augmented speech data into sub-networks.
• Weighted fusion of feature maps from sub-networks for final emotional representation.
• Adversarial training with a discriminator to estimate Wasserstein distance for speaker-invariant features.

Main Results:

• The proposed model demonstrated superior performance on the IEMOCAP corpus.
• Performance gains ranged from 2-9% compared to existing SER algorithms.
• Adversarial training successfully generalized emotion representations across speakers.

Conclusions:

• Effective data augmentation significantly improves SER model performance.
• The proposed multi-channel, adversarial training approach yields a robust and generalized SER system.
• The method is effective in overcoming limitations of labeled data scarcity in SER.