A novel speech emotion recognition method based on feature construction and ensemble learning.

Area of Science:

• Human-Computer Interaction (HCI)
• Machine Learning
• Speech Processing

Background:

• Speech emotion recognition (SER) is crucial in HCI.
• Limited speech emotion datasets pose a significant challenge for model training.
• Existing methods may not fully leverage feature engineering and ensemble techniques.

Purpose of the Study:

• To propose a novel SER method addressing data scarcity.
• To enhance SER accuracy through effective feature construction and ensemble learning.
• To develop a robust ensemble model integrating multiple weak learners.

Main Methods:

• Acoustic features were extracted and combined into diverse feature sets.
• A novel feature selection method, L-SFS (LightGBM and Sequential Forward Selection), was developed.
• An ensemble model, Sklex, was created using weighted averaging of SVM, KNN, XGBoost, and LightGBM, with weights learned via softmax regression.

Main Results:

• The proposed L-SFS method effectively selects relevant features.
• The Sklex ensemble model demonstrated superior and stable performance.
• The method achieved good speech emotion recognition accuracy, validating its effectiveness.

Conclusions:

• Feature construction is vital for improving SER performance.
• Ensemble learning offers significant advantages in terms of accuracy and stability for SER.
• The proposed method provides a promising solution for SER with limited data.