Depression Detection from Audio (DAIC-WOZ)

An AI-powered screening tool that analyzes audio clips to detect markers associated with depression. This project uses a Bidirectional Long Short-Term Memory (BiLSTM) neural network trained on the DAIC-WOZ dataset.

Key Features

• High-Fidelity Audio Analysis: Uses 40 MFCCs plus Delta and Delta-Delta features (120 total features per frame).
• Proactive Pre-processing: Automatic silence trimming and per-sample z-normalization.
• Deep Learning Architecture: Stacked Bidirectional LSTMs with Batch Normalization for temporal feature learning.
• Interactive UI: Clean Streamlit interface for uploading audio and viewing real-time predictions.

Setup & Installation

1. Clone the repository (or navigate to the directory).
2. Create and activate a virtual environment:

   python -m venv venv
   # On Windows:
   venv\Scripts\activate
   # On Mac/Linux:
   source venv/bin/activate

3. Install dependencies:

   pip install streamlit tensorflow librosa numpy pandas scikit-learn

How to Run

1. Train/Download the model:
   • Ensure you have best_depression_model.keras and model_config.json inside the models/ folder.
2. Start the application:

   streamlit run app.py

3. Analyze: Upload a .wav file or use the built-in test audio.

Model Performance

The current model utilizes an optimized threshold of 0.34, achieving:

• Accuracy: ~74%
• Recall (Depressed): ~86%
• F1-Score (Depressed): ~0.76

Disclaimer

This tool is for research purposes only and is not a clinical diagnostic instrument. It is designed to demonstrate the potential of vocal biomarkers in mental health screening. Always consult a medical professional for clinical diagnosis.