A Combined CNN Architecture for Speech Emotion Recognition.

Area of Science:

• Artificial Intelligence
• Speech Processing
• Human-Computer Interaction

Background:

• Emotion recognition from speech is crucial for Human-Computer Interaction (HCI).
• Existing deep learning methods face challenges with data quantity, diversity, and feature selection standards.
• Designing effective neural network architectures for speech emotion recognition remains complex.

Purpose of the Study:

• To address limitations in current speech emotion recognition techniques using deep learning.
• To propose a comprehensive approach involving data preprocessing, feature selection, and a novel neural network architecture.
• To develop and utilize a unified dataset (EmoDSc) for robust emotion recognition.

Main Methods:

• Construction of a unified dataset (EmoDSc) by combining existing speech emotion databases.
• Investigation of the synergy between spectral features and spectrogram images for emotion recognition.
• Development of a hybrid neural network architecture integrating 1D Convolutional Neural Network (CNN1D), 2D CNN (CNN2D), and Multilayer Perceptron (MLP) to fuse spectral and image-based features.

Main Results:

• Individual analysis showed weighted accuracies of 89% for spectral features and 90% for spectrogram images.
• The proposed hybrid model, utilizing the EmoDSc dataset, achieved a remarkable weighted accuracy of 96%.
• The fused approach significantly outperformed models relying on isolated feature types.

Conclusions:

• The proposed deep learning approach, combining spectral features and spectrogram images via a hybrid CNN-MLP architecture, significantly advances speech emotion recognition.
• The unified EmoDSc dataset provides a valuable resource for training and evaluating speech emotion recognition models.
• This study offers a robust solution to improve the accuracy and reliability of emotion recognition in human-computer interaction systems.