Effects of Data Augmentations on Speech Emotion Recognition

Area of Science:

• Speech Processing
• Machine Learning
• Affective Computing

Background:

• Data augmentation is increasingly used in speech emotion recognition (SER).
• However, increasing data volume does not always guarantee improved model performance, suggesting a potential trade-off.

Purpose of the Study:

• To investigate the impact of data augmentation on SER performance.
• To identify optimal data augmentation techniques and quantities for different conditions.

Main Methods:

• Experiments were conducted on the Japanese Twitter-based emotional speech and IEMOCAP datasets.
• Evaluated various data augmentation strategies, including glottal source extraction and silence removal.

Main Results:

• For speaker-independent SER, two augmentation techniques (glottal source extraction and silence removal) yielded the best results.
• For text-independent SER, increased data augmentation generally improved performance.

Conclusions:

• A trade-off exists between the amount of data augmentation and SER performance.
• Selecting appropriate data augmentation methods is crucial for specific SER conditions.