Speech emotion recognition based on a stacked autoencoders optimized by PSO based grass fibrous root optimization.

Area of Science:

• Artificial Intelligence
• Signal Processing
• Computational Linguistics

Background:

• Speech emotion recognition (SER) is challenging due to the subjective nature of human emotions.
• Accurate SER has applications in healthcare, human-computer interaction, and social robotics.
• Existing SER models face limitations in effectively capturing nuanced emotional cues.

Purpose of the Study:

• To develop an innovative and efficient speech emotion recognition system.
• To improve the accuracy of identifying emotional states from speech signals.
• To integrate deep learning with metaheuristic optimization for enhanced SER performance.

Main Methods:

• A stacked autoencoder (SAE) was employed as the core deep learning model.
• The SAE's performance was fine-tuned using a hybrid metaheuristic algorithm combining Particle Swarm Optimization (PSO) and Grass Fibrous Root Optimization (GFRO).
• Spectral and pitch features, including spectral crest, entropy, flux, and harmonic ratio, were extracted from speech signals.

Main Results:

• The proposed hybrid deep learning model demonstrated high accuracy in speech emotion recognition.
• Performance was evaluated on a standard emotion recognition dataset.
• The model outperformed several state-of-the-art methods, including CNN, SVM, DL, CNN/INCA, and VGG-16.

Conclusions:

• The integration of SAE with PSO-GFRO offers a powerful approach for effective speech emotion recognition.
• The method successfully extracts relevant features for accurate emotion identification.
• This research contributes a significant advancement to the field of SER, with potential for real-world applications.