Speech Emotion Recognition Using Attention Model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Healthcare Technology

Background:

• Speech emotion recognition is crucial for public health and advancing healthcare technology.
• Deep learning models and novel acoustic/temporal features have improved speech emotion recognition systems.
• Existing research lacks a comprehensive analysis of feature combinations for optimal performance.

Purpose of the Study:

• To propose a self-attention-based deep learning model for enhanced speech emotion recognition.
• To identify the most effective spectral and rhythmic features for emotion detection.
• To evaluate the model's performance on a combined dataset of multiple emotion corpora.

Main Methods:

• A novel deep learning architecture combining a 2D Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) network with self-attention was developed.
• Extensive experiments were conducted using various combinations of spectral and rhythmic features.
• Mel Frequency Cepstral Coefficients (MFCCs) were selected as the optimal feature set.
• The model was trained and tested on a custom dataset comprising RAVDESS, SAVEE, and TESS datasets, detecting eight emotional states.

Main Results:

• The proposed self-attention-based deep learning model achieved an average test accuracy of 90%.
• Mel Frequency Cepstral Coefficients (MFCCs) demonstrated superior performance compared to other feature combinations.
• The model successfully detected eight distinct emotional states: happy, sad, angry, surprise, disgust, calm, fearful, and neutral.

Conclusions:

• The developed attention-based deep learning model significantly improves speech emotion recognition accuracy.
• This model, utilizing MFCCs, shows great potential for enhancing automated mental health monitoring systems.
• The findings contribute to the advancement of affective computing and personalized healthcare solutions.