Depression Detection from Audio

This repository contains experiments for detecting depression from speech using deep learning models. The project leverages the autrainer toolkit, a modular framework built on PyTorch and Hydra for reproducible computer audition research.

The primary goal is to classify patients as depressive or non-depressive based on features extracted from their speech. Two main architectural approaches are explored:

1. CNN10 Model: A Convolutional Neural Network trained directly on log-Mel spectrograms for binary classification.
2. CNN-LSTM Model: Uses a pre-trained CNN10 as a feature extractor, followed by an LSTM with an attention mechanism to model temporal sequences.

Table of Contents

• Project Structure
• 1. Setup and Installation
  • 1.1. Clone the Repository
  • 1.2. Create a Virtual Environment
  • 1.3. Install Dependencies
• 2. Environment Configuration
  • 2.1. Create Environment File
• 3. Data Preparation
  • 3.1. Fetch Data
  • 3.2. Preprocess Data
• 4. Running Experiments
  • Experiment 1: CNN10 Binary Classification
  • Experiment 2: CNN-LSTM Training
    • Extract CNN Features
    • Train the LSTM Model
• 5. Post-processing and Analyzing Results
• 6. Inference
• 7. Manual Evaluation

Project Structure

The repository is organized to support reproducible experiments with autrainer and Hydra.

/
├── conf/              # Hydra/autrainer configuration files for datasets, models, etc.
├── data/              # Processed data (e.g., features, spectrograms).
├── data_raw/          # Raw data downloaded by fetch commands.
├── results/           # Experiment outputs, logs, and trained models.
├── requirements.txt   # Project dependencies.
├── postprocess.sh     # Helper script to analyze results.
└── README.md          # This file.

1. Setup and Installation

Follow these steps to set up the environment and install the necessary dependencies.

1.1. Clone the Repository

git clone <repository-url>
cd <repository-name>

1.2. Create a Virtual Environment

It is highly recommended to use a virtual environment.

python -m venv venv
source venv/bin/activate

1.3. Install Dependencies

Install all required packages, including autrainer and its dependencies.

pip install -r requirements.txt

This project may require openSMILE for certain feature extraction configurations. If needed, install it separately and ensure it's available in your system's PATH. You can install the optional opensmile dependency for autrainer with:

pip install autrainer[opensmile]

2. Environment Configuration

Before running the experiments, you need to configure the dataset URL in an environment file.

2.1. Create Environment File

Copy the example environment file and fill in the Extended DAIC-WOZ dataset link:

cp .env.example .env

Then edit the .env file and replace <LINK TO DATASET> with the actual URL to the Extended DAIC-WOZ dataset. You can reference .env.example to see the expected format:

BASE_URL=<LINK TO DATASET>

3. Data Preparation

The experiments rely on the Extended DAIC-WOZ dataset. The following commands will download and preprocess the data into the required format.

3.1. Fetch Data

This command downloads the necessary datasets and pre-trained model weights specified in the configuration files.

autrainer fetch

3.2. Preprocess Data

This command processes the raw audio files into the features required for training (e.g., log-Mel spectrograms). The configurations in the conf/ directory are set up to handle this automatically.

autrainer preprocess

4. Running Experiments

All experiments are managed by autrainer and can be launched from the command line. The results, including logs, model checkpoints, and plots, will be saved in the results/ directory.

Experiment 1: CNN10 Binary Classification

This experiment trains a CNN10 model directly on log-Mel spectrograms for end-to-end depression classification.

Run the baseline training with a fixed dataset split:

autrainer train -cn config-fixed

Experiment 2: CNN-LSTM Training

This workflow involves two stages: extracting features with the CNN, and then training the LSTM on those features.

Extract CNN Features

First, run the feature extraction script. This uses the pre-trained CNN10 model to generate feature sequences for each patient and saves them in data/ExtendedDAIC-lstm/features/.

python extract_cnn_features.py \
    --data_path data/ExtendedDAIC-16k \
    --output_path data/ExtendedDAIC-lstm \
    --model_path model.pt

Train the LSTM Model

Once the features are extracted, train the LSTM model. This script runs a standalone training process and logs results to Weights & Biases.

python lstm_standalone.py

5. Post-processing and Analyzing Results

autrainer provides tools to analyze the results of your experiments, especially for grid searches.

To summarize the results of an experiment and aggregate across different seeds:

# Replace <experiment_id> with the one from your config (e.g., cnn10-fixed)
autrainer postprocess results/<experiment_id> --aggregate seed

This will generate summary CSVs and plots in the results/<experiment_id>/summary/ directory. You can also use the helper script:

./postprocess.sh <experiment_id>

6. Inference

To run inference on new audio files using a trained model, use the autrainer inference command.

You need to point to a trained model directory, an input directory with audio files, and an output directory.

# Example command for a trained model from the 'cnn10-fixed' experiment
# Note: The path to the specific run may vary.
autrainer inference \
    results/cnn10-fixed/training/ExtendedDAIC-16k-fixed_CNN10-binary_Adam_0.0001_32_epoch_50_None_None_42/ \
    /path/to/your/input_audio/ \
    /path/to/your/output_predictions/ \
    --preprocess-cfg log_mel_16k \
    --device cuda:0

7. Manual Evaluation

The repository includes a script to manually calculate detailed metrics from prediction files. After generating depression_predictions.csv and snippet_predictions.csv (e.g., using predict_depression.py), you can get a full evaluation report.

python calculate_metrics.py