A Comparison of Machine Learning Algorithms and Feature Sets for Automatic Vocal Emotion Recognition in Speech.

Area of Science:

• Speech processing
• Machine learning
• Affective computing

Background:

• Vocal emotion recognition (VER), also known as speech emotion recognition (SER), is difficult for both humans and machines.
• Advancements in VER algorithms are crucial for applications in clinical diagnosis, social interaction research, and Human-Computer Interaction (HCI).
• Optimal low-level descriptors and machine learning classifiers for VER are not yet established.

Purpose of the Study:

• To compare the performance of various machine learning algorithms using different feature sets for vocal emotion recognition.
• To identify the most effective combination of feature sets and classifiers for accurate emotion classification in speech.

Main Methods:

• Seven machine learning algorithms (MLP, J48 DT, SMO, RF, KNN, LOG, MLR) were evaluated.
• The Berlin Database of Emotional Speech was used with 10-fold cross-validation.
• Four openSMILE feature sets (IS-09, emobase, GeMAPS, eGeMAPS) were employed.

Main Results:

• Support Vector Machine with Sequential Minimal Optimization (SMO), Multilayer Perceptron Neural Network (MLP), and Simple Logistic Regression (LOG) demonstrated superior performance.
• SMO achieved the highest accuracy at 87.85%, followed by MLP (84.00%) and LOG (83.74%).
• The emobase feature set yielded the best overall results across the evaluated algorithms.

Conclusions:

• The study highlights the effectiveness of specific machine learning algorithms and feature sets for VER.
• Findings provide valuable insights for developing efficient VER systems in clinical diagnosis, intervention, and HCI.
• The emobase feature set is recommended for its strong performance in speech emotion classification.