Modeling Depression as a Gradient Condition Using Acoustic Features of Mandarin Speech

Area of Science:

• Speech analysis
• Psychiatry
• Computational linguistics

Background:

• Depression is common and often missed.
• Speech analysis for depression shows promise, but research is limited to non-tonal languages and binary classification.
• Mandarin, a tonal language, presents unique challenges and opportunities for speech-based depression detection.

Purpose of the Study:

• To investigate if acoustic features in spontaneous Mandarin speech indicate depression as a category or a continuous dimension.
• To explore the utility of speech analysis in understanding the nuances of depression.

Main Methods:

• Analysis of a Mandarin speech corpus with depression severity scores.
• Extraction of acoustic features using the extended Geneva Minimalistic Acoustic Parameter Set.
• Application of random forest classification, linear mixed-effects modeling, and unsupervised clustering.

Main Results:

• Random forest achieved 72.3% accuracy in distinguishing depressed from non-depressed speech.
• Key acoustic features identified include spectral, pitch, voice quality, cepstral, and intensity measures.
• Fundamental frequency positively correlated with depression severity, supporting a dimensional view.

Conclusions:

• Findings suggest acoustic features in Mandarin speech align better with depression severity than diagnostic groups.
• Speech-based measures show potential for continuous depression monitoring.
• Further longitudinal studies are needed to confirm these findings.