A generalizable speech emotion recognition model reveals depression and remission

Area of Science:

• Speech analysis
• Machine learning
• Mental health

Background:

• Affective disorders, such as major depressive disorder (MDD), are linked to atypical voice patterns.
• Automated voice analysis for clinical use faces challenges with small sample sizes and generalizability.
• Transfer learning offers a promising approach to overcome these limitations.

Purpose of the Study:

• To investigate a generalizable method for evaluating depression and remission using voice analysis via transfer learning.
• To train machine learning models on non-clinical datasets and test their efficacy on clinical data in a different language.

Main Methods:

• A Mixture of Experts machine learning model was trained on German and US English emotional speech corpora.
• The model's ability to classify depression was tested on Danish-speaking individuals: healthy controls (N=42), first-episode MDD patients (N=40), and patients in remission (N=25).
• Evaluation included raw, de-noised, and speaker-diarized clinical interview recordings.

Main Results:

• The model distinguished between healthy controls and depressed patients with an AUC of 0.71.
• Speech from patients in remission was not distinguishable from that of the control group.
• Predictions were stable, suggesting 20-30 seconds of speech may suffice for screening; background noise significantly impacted results.

Conclusions:

• A generalizable speech emotion recognition model can effectively track changes in depressive states before and after remission in MDD patients.
• Data collection settings and cleaning are critical for the clinical application of automated voice analysis.