Acoustics of depression

The sound of speech reflects the speaker's mood in a way that may enable objective measurement of depression from speech audio recordings. As previous studies relied on volunteers without a diagnosis or selected subsets of acoustic features, it is unknown which acoustic features measure clinically relevant depression.

We examined a comprehensive set of 77 system, source, spectral, and prosodic features of speech in 239 adults, including 147 with major depressive disorder, who spoke following neutral, positive, and negative prompts. We randomly divided the participants into an analysis set of 176 and a test set of 63 individuals. In the training set, we quantified the relationships between acoustic features and depression severity on the Montgomery Åsberg Depression Rating Scale (MADRS). We built XGBoost machine learning models to differentiate depression severity levels (none, mild, moderate, severe) and validated them in the unseen test set.

Multiple acoustic features discriminated depression severity levels with prosodic features contributing the most information. Speech energy velocity, intensity variation, voiced rate, pause duration (prosodic features) and the third formant variance (a system feature) predicted depression severity consistently across prompts. Intensity, energy, pause duration, and harmonic difference contributed the most to XGBoost models that discriminated severity levels in the test set.

The analysis is limited to speech in English.

Acoustic, in particular prosodic, features can discriminate clinically relevant levels of depression among diagnosed patients. The robustness of these predictors across emotional contexts suggests a potential for application in a variety of settings.