Multi-Head Attention-Based Long Short-Term Memory for Depression Detection From Speech.

Area of Science:

• Computer Science
• Artificial Intelligence
• Psychology

Background:

• Depression detection is crucial for mental health, yet current methods often underutilize speech's inherent information.
• Existing research frequently relies on multi-modal features, overlooking the potential of detailed speech analysis.

Purpose of the Study:

• To propose a novel deep learning model for enhanced depression detection using speech.
• To leverage time-dimension attention mechanisms within a Long Short-Term Memory (LSTM) network to focus on salient speech characteristics.

Main Methods:

• Extraction of frame-level speech features to preserve temporal dynamics.
• Development of a multi-head time-dimension attention mechanism applied to LSTM outputs.
• Utilizing modified feature sets for improved input to the LSTM layer.

Main Results:

• The proposed multi-head attention-LSTM model demonstrated improved depression detection performance.
• Performance gains of 2.3% on the DAIC-WOZ corpus and 10.3% on the MODMA corpus were observed compared to a standard LSTM model.
• The attention mechanism successfully identified key temporal information indicative of depression in speech.

Conclusions:

• The multi-head time-dimension attention-based LSTM model offers a more effective approach to depression detection from speech.
• Focusing on emotionally salient speech regions and temporal patterns enhances diagnostic accuracy.
• This method provides a promising direction for developing objective, speech-based mental health assessment tools.