Cross-corpus speech emotion recognition with transformers: Leveraging handcrafted features and data augmentation.

Area of Science:

• Data Science
• Artificial Intelligence
• Speech Processing

Background:

• Speech Emotion Recognition (SER) is crucial for applications like psychological assessment and human-computer interaction.
• Traditional SER methods using manually engineered features struggle with complex data variations, noise, and cross-dataset inconsistencies.
• Existing techniques often show reduced classification accuracy in diverse and challenging acoustic environments.

Purpose of the Study:

• To develop and evaluate a novel SER framework to improve emotion classification accuracy.
• To address limitations of manual feature engineering in handling diverse speech patterns and noisy data.
• To create a robust cross-corpus model for enhanced SER performance across different datasets.

Main Methods:

• Applied signal preprocessing and data augmentation to enhance raw speech signals.
• Derived 18 informative features and employed feature selection for a discriminative feature set.
• Utilized the SAVEE, RAVDESS, and EMO-DB datasets for training and testing, including a cross-corpus model.

Main Results:

• The proposed SER framework achieved superior performance compared to existing methods.
• High accuracy rates were obtained: 95% on SAVEE, 94% on RAVDESS, 97% on EMO-DB, and 97% on the cross-corpus model.
• The framework demonstrated effectiveness in handling diverse emotional expressions and data complexities.

Conclusions:

• The novel SER framework significantly advances the state-of-the-art in speech emotion recognition.
• The integrated approach of preprocessing, augmentation, and feature selection yields robust performance.
• The study highlights the potential of the proposed model for real-world SER applications.