Robust Multi-Scenario Speech-Based Emotion Recognition System.

Area of Science:

• Speech processing
• Affective computing
• Machine learning

Background:

• Human emotions are daily experiences often conveyed through speech.
• Existing acoustic emotional speech databases frequently lack real-world conditions, using acted emotions and no noise.
• Limited data patterns in current databases lead to generalization issues and overfitting in emotion recognition systems.

Purpose of the Study:

• To investigate the impact of environmental factors like noise and reverberation on emotional speech recognition (ESR) system performance.
• To propose a robust multi-scenario system and a virtual enlargement method to improve ESR in realistic conditions.

Main Methods:

• Experiments were conducted using logistic regression to simulate various environmental conditions.
• Scenarios included different levels of Gaussian white noise, real-world noise, and reverberation.
• A virtual enlargement method was employed to enhance the robustness of the proposed system.

Main Results:

• System performance deteriorated significantly under all tested environmental conditions.
• Error probability increased from 25.57% to a worst-case 79.13% due to simulated noise and reverberation.
• The proposed robust system achieved an average error probability of 34.57%, comparable to ideal conditions (31.55%).

Conclusions:

• Simulated emotional speech databases do not adequately represent real-world scenarios, impacting system reliability.
• Environmental factors like noise and reverberation are critical challenges for accurate speech-based emotion recognition.
• The developed robust multi-scenario system demonstrates improved performance in handling diverse and challenging acoustic environments.