Fusing traditionally extracted features with deep learned features from the speech spectrogram for anger and stress...

Area of Science:

• Computer Science
• Speech Processing
• Artificial Intelligence

Background:

• Negative emotions like stress and anger impact mental and physical health.
• Automated systems are crucial for early detection of emotional health issues.
• Speech signals contain valuable information for emotion recognition.

Purpose of the Study:

• To propose a computationally efficient convolutional neural network (CNN) for detecting stress and anger.
• To enhance emotion recognition accuracy by combining handcrafted and deep learned features from speech spectrograms.
• To develop an automated system for monitoring mental states and identifying early signs of emotional distress.

Main Methods:

• Utilized a convolutional neural network (CNN) architecture.
• Extracted features from speech spectrograms, including handcrafted and deep learned features.
• Evaluated the proposed method on three diverse speech emotion datasets: TESS, RAVDESS, and EMO-DB.

Main Results:

• The CNN model demonstrated high categorical accuracy across all tested datasets.
• Achieved training accuracies of 93.7% (TESS), 97.5% (EMO-DB), and 96.7% (RAVDESS).
• Achieved validation accuracies of 95.6% (TESS), 95.6% (EMO-DB), and 96.7% (RAVDESS), with reduced cross-entropy loss.

Conclusions:

• The proposed CNN approach effectively detects stress and anger from speech signals.
• Combining feature sets significantly boosts recognition accuracy and reduces computational load.
• The method shows promise for developing automated systems for mental state monitoring and early emotional health issue detection.