Feature selection enhancement and feature space visualization for speech-based emotion recognition.

Area of Science:

• Computer Science
• Signal Processing
• Artificial Intelligence

Background:

• Robust speech emotion recognition (SER) is crucial for human-computer interaction.
• The quality of extracted speech features directly impacts SER system performance.
• Existing feature extraction methods may not capture all relevant emotional nuances.

Purpose of the Study:

• To propose and evaluate a novel speech feature enhancement strategy for improved SER.
• To investigate the effectiveness of principal component analysis (PCA) and feature fusion on established feature sets.
• To validate the proposed method on diverse, multilingual emotional speech datasets.

Main Methods:

• Utilized the INTERSPEECH 2010 challenge feature set.
• Applied principal component analysis (PCA) to identified feature subsets.
• Implemented horizontal feature fusion.
• Visualized feature space using t-distributed stochastic neighbor embedding (t-SNE).
• Evaluated performance on the Berlin Emotional Speech Dataset (EMO-DB) and Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS).

Main Results:

• Achieved an average recognition gain of 11.5% for six out of seven emotions on the EMO-DB dataset (German).
• Achieved an average recognition gain of 13.8% for seven out of eight emotions on the RAVDESS dataset (English).
• Demonstrated superior performance compared to baseline methods and state-of-the-art approaches.

Conclusions:

• The proposed speech feature enhancement strategy significantly improves SER accuracy.
• PCA-based feature subset analysis and fusion are effective for enhancing emotional discriminability.
• The method shows strong generalization capabilities across different languages and emotional speech corpora.