Speech based suicide risk recognition for crisis intervention hotlines using explainable multi-task learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• Crisis intervention hotlines face challenges with low connectivity and delayed responses.
• Integrating speech signals and deep learning can enhance crisis assessment and intervention effectiveness.

Purpose of the Study:

• To develop and validate a novel speech crisis recognition method for suicide risk assessment.
• To explore gender differences and speech feature variability in crisis calls.

Main Methods:

• Constructed a crisis intervention hotline suicide risk speech dataset labeled using the Modified Suicide Risk Scale.
• Employed a data-theoretically dual-driven, gender-assisted speech crisis recognition method based on multi-tasking and deep learning.
• Utilized five-fold cross-validation for model evaluation.

Main Results:

• Identified gender differences in speech duration during crisis calls (males spoke more than females).
• Found significant variations in emotional intensity, speech rate, and texture among crisis callers.
• The proposed method achieved a 96% F1 score, outperforming existing methods.

Conclusions:

• The developed model demonstrates high effectiveness in speech crisis recognition.
• Statistical analysis and integration of data with theoretical knowledge improve model interpretability and effectiveness.
• Future work should consider larger sample sizes and multimodal data integration.