An Online Attachment Style Recognition System Based on Voice and Machine Learning.

Area of Science:

• Psychology
• Computer Science
• Health Informatics

Background:

• Insecure attachment styles are linked to increased risks of mental and physical health issues.
• Objective assessment of attachment styles is crucial for early intervention and risk management.

Purpose of the Study:

• To develop a machine learning model for distinguishing secure from insecure attachment styles using voice recordings.
• To investigate the role of acoustic features and gender differences in attachment style recognition.

Main Methods:

• 199 participants provided voice recordings in response to attachment-triggering questions.
• Acoustic features were extracted using the eGeMAPS set, with recursive feature elimination for selection.
• Supervised machine learning models were trained using gender-independent and gender-dependent approaches.

Main Results:

• The gender-independent model achieved 58.88% test accuracy.
• Gender-dependent models showed higher accuracy: 63.88% for women and 83.63% for men.
• Results highlight significant gender influence on voice-based attachment style recognition.

Conclusions:

• Acoustic properties hold potential for remote, objective assessment of attachment styles.
• Considering gender separately improves model performance, suggesting tailored approaches for health risk identification.
• This research supports developing large-scale mobile screening systems for attachment-related health risks.