A Deep Learning Method Using Gender-Specific Features for Emotion Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Speech Processing

Background:

• Speech analysis offers insights into mental states and aids human-computer interaction.
• Speech emotion recognition (SER) is valuable for mental illness diagnosis but is hindered by gender-based acoustic variations.
• Existing SER models often show reduced accuracy due to overlooking gender-specific speech characteristics.

Purpose of the Study:

• To enhance speech emotion recognition accuracy by developing gender-specific models.
• To investigate the impact of gender on acoustic features relevant to emotion recognition.
• To propose a novel SER method incorporating gender classification and optimized feature selection.

Main Methods:

• Gender classification of speech using a Multilayer Perceptron (MLP).
• Analysis of acoustic feature influence weights for male and female speech separately.
• Establishment of distinct, optimal feature sets for male and female SER.
• Training and testing of Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) networks on gender-specific feature sets.

Main Results:

• The proposed gender-specific SER models demonstrated improved average recognition accuracy.
• Gender classification effectively enabled the selection of optimal features for each gender.
• CNN and BiLSTM models trained on tailored feature sets outperformed gender-mixed recognition models.

Conclusions:

• Gender-specific feature selection significantly enhances speech emotion recognition performance.
• A gender classification-based approach is a viable strategy for improving SER systems.
• This method offers a more accurate and reliable approach to SER for mental health applications.