INFERRING CLINICAL DEPRESSION FROM SPEECH AND SPOKEN UTTERANCES

Area of Science:

• Computational linguistics
• Psychiatry
• Speech processing

Background:

• Depression detection from speech is gaining interest for objective clinical assessment.
• Speech cues for depression exist in content and prosody.
• Traditional n-gram models are limited by data constraints.

Purpose of the Study:

• To investigate depression detection using speech processing and machine learning.
• To explore alternative feature spaces beyond n-grams, specifically valence and arousal.
• To evaluate novel harmonic model-based prosody features against standard methods.

Main Methods:

• Utilized word representations in valence-arousal feature space.
• Employed standard openSMILE for prosody extraction.
• Developed and compared novel harmonic model features for prosody.
• Integrated content and prosody features for improved detection.

Main Results:

• Harmonic model features enhanced depression detection performance.
• Combined features achieved approximately 74% accuracy.
• The developed approach shows promise for screening applications.

Conclusions:

• Speech analysis, particularly using harmonic models for prosody, is effective for depression detection.
• The integration of content and prosody features offers a robust approach.
• This method provides a valuable tool for supplementary clinical information and screening.