Beyond acoustic features: Incorporating linguistic variables in automatic speech analysis for depression detection.

Area of Science:

• Speech analysis
• Computational linguistics
• Clinical psychology

Background:

• Automatic speech analysis (ASA) predominantly uses acoustic features, neglecting linguistic markers.
• Linguistic markers are underexplored in non-English speaking, clinically diagnosed populations.

Purpose of the Study:

• To evaluate the integration of acoustic and linguistic markers for depression detection.
• To assess the predictive value of these markers in a Spanish-speaking clinical sample.

Main Methods:

• 151 participants (80 with MDD/PDD, 71 controls) provided speech via open-ended questions.
• Extracted linguistic and acoustic features (prosodic, cepstral, spectral, Teager Energy Operator).
• Employed logistic regression and machine learning models for classification.

Main Results:

• Linguistic features (e.g., verb/noun usage, vocabulary size) were strong depression predictors.
• The linguistic model (AUC=0.86) outperformed the acoustic model (AUC=0.79).
• Ensemble model achieved comparable performance (AUC=0.86) with high accuracy (0.84) and specificity (0.93).

Conclusions:

• Integrating linguistic features into ASA enhances depression detection.
• Speech-based assessments hold potential for early depression screening in primary care settings.