Speech Emotion Recognition Based on Selective Interpolation Synthetic Minority Over-Sampling Technique in Small...

Area of Science:

• Artificial Intelligence
• Speech Processing
• Machine Learning

Background:

• Speech emotion recognition (SER) faces challenges with imbalanced data and feature redundancy across diverse applications.
• Existing SER models are often tailored to specific sample conditions, limiting generalizability.
• A robust SER model for small sample environments is needed.

Purpose of the Study:

• To propose an effective speech emotion recognition model for small sample environments.
• To address data imbalance and feature redundancy in SER.
• To improve the accuracy and robustness of SER systems.

Main Methods:

• Developed a data imbalance processing method using Selective Interpolation Synthetic Minority Over-sampling Technique (SISMOTE).
• Implemented a feature selection method combining variance analysis and Gradient Boosting Decision Tree (GBDT) to remove redundant features.
• Evaluated the proposed model on CASIA, Emo-DB, and SAVEE speech emotion databases.

Main Results:

• Achieved high speaker-dependent speech emotion recognition accuracies: 90.28% on CASIA, 75.00% on SAVEE, and 85.82% on Emo-DB.
• Demonstrated superior performance compared to several state-of-the-art methods.
• Effectively reduced the impact of sample imbalance and excluded poorly representative features.

Conclusions:

• The proposed SISMOTE and GBDT-based feature selection method significantly enhances SER performance in small sample scenarios.
• The model offers a robust and accurate solution for emotion recognition from speech, outperforming existing approaches.
• This work provides a valuable contribution to the field of affective computing and human-computer interaction.