Speech emotion classification using attention based network and regularized feature selection.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• Speech emotion classification (SEC) is crucial for Human-Computer Interaction (HCI) and affective computing.
• Existing deep neural network (DNN) models face challenges with multi-lingual data and other factors affecting accurate emotion recognition.
• Attention mechanisms show promise in sequence-based and time-series tasks.

Purpose of the Study:

• To propose an improved speech emotion classification model using an attention-based network.
• To integrate a pre-trained convolutional neural network (CNN) with regularized neighbourhood component analysis (RNCA) for feature selection.
• To enhance the accuracy and performance of emotion recognition from speech.

Main Methods:

• Developed an attention-based deep convolutional neural network (DCNN).
• Employed regularized neighbourhood component analysis (RNCA) for feature selection.
• Conducted experiments using Support Vector Machine (SVM), Multilayer Perceptron (MLP), and Random Forest (RF) classifiers on the TESS dataset.

Main Results:

• The proposed Attention-based DCNN+RNCA+RF model achieved 97.8% classification accuracy.
• Demonstrated a 3.27% performance improvement over state-of-the-art SEC approaches.
• The attention mechanism and feature selection methods aligned with human emotion perception patterns.

Conclusions:

• The attention-based DCNN with RNCA feature selection offers superior performance for speech emotion classification.
• The model's effectiveness highlights the potential of attention mechanisms and robust feature selection in affective computing.
• This approach provides a more consistent and accurate method for recognizing emotions from auditory speech.