Speech Emotion Recognition Incorporating Relative Difficulty and Labeling Reliability.

Area of Science:

• Computer Science
• Artificial Intelligence
• Speech Processing

Background:

• Speech emotion recognition (SER) models often struggle with unseen data due to varying emotional expressions.
• Existing methods like regularization or metric losses aim to improve SER robustness.
• Challenges arise from emotional context and annotator reliability affecting training data.

Purpose of the Study:

• To develop a speech emotion recognition (SER) method robust to unseen corpora.
• To incorporate training sample difficulty and labeling reliability into SER model training.
• To improve SER performance on diverse datasets beyond the training distribution.

Main Methods:

• Proposed a novel loss function inspired by Proxy-Anchor loss, prioritizing harder-to-classify samples.
• Introduced label smoothing for samples misclassified by a pre-trained SER model to mitigate unreliable labels.
• Integrated relative sample difficulty and label reliability into the training process.

Main Results:

• The proposed SER method demonstrated improved performance on unseen corpora.
• The novel loss function effectively guided the model by focusing on challenging samples.
• Label smoothing on misclassified data further enhanced the model's generalization capabilities.

Conclusions:

• The proposed SER approach enhances model robustness against variations in unseen speech emotion data.
• Incorporating sample difficulty and label reliability is crucial for effective SER.
• The combination of the novel loss function and label smoothing offers a promising direction for future SER research.