Depression Estimation Using Audio with Machine Learning Models

Project Overview

This project explores the automatic estimation of depression levels based on acoustic characteristics of speech, without analyzing the content of the spoken words. It uses machine learning and deep learning techniques to predict PHQ-8 scores (0-24) from audio recordings of clinical interviews.

The project was developed by Veselin Roganović (SV 36/2022) as part of the Soft Computing course at the Faculty of Technical Sciences, Novi Sad.

Dataset

• Name: DAICWOZ
• Content: Audio recordings of clinical interviews.
• Targets: PHQ-8 depression scores.
• Sample Size: 140 participants.
• Data Split: 80% Training, 10% Validation, 10% Test.

Preprocessing

• Resampling: All audio is resampled to 16kHz.
• Normalization: Amplitude normalization is applied.
• Feature Extraction: MFCC (Mel-frequency cepstral coefficients) are extracted using a sliding window approach to segment long interviews and augment the data.
  • Window size: ~10-42s (depending on model configuration)
  • Hop length: ~5-18s

Methodology

Two primary approaches were implemented and compared:

1. SVR (Support Vector Regression)

• Input: Flattened MFCC matrices from time windows.
• Model: SVR with an RBF kernel.
• Hyperparameters: C=1, epsilon=0.01, gamma='scale'.
• Scaler: StandardScaler from Scikit-Learn.

2. CNN + LSTM

• Architecture:
  • CNN: 2D convolutional layers to extract spatial features from MFCC representations, including BatchNormalization and Dropout.
  • LSTM: 2 layers with 64 hidden units to capture temporal dependencies across sequences.
  • Output: A fully connected layer for final regression.
• Loss Function: A combined loss of MAE and Pearson Correlation Loss to optimize both absolute error and trend prediction.

Results

Evaluation metrics include Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Pearson Correlation coefficient.

Validation Results

Model	MAE	RMSE	Pearson
SVR	4.29	4.98	0.52
CNN+LSTM	3.99	5.02	0.45

Test Results

Model	MAE	RMSE	Pearson
SVR	4.37	6.17	0.27
CNN+LSTM	5.51	7.20	0.05

Conclusion: SVR outperformed the deep learning approach on the test set, suggesting it is more robust for small-scale clinical datasets where deep models are prone to overfitting.

Technologies Used

• Python
• PyTorch (Deep Learning)
• Scikit-Learn (SVR & Preprocessing)
• Librosa (Audio Processing)
• Pandas/NumPy (Data Management)

How to Run

1. Prepare Environment: Install the required dependencies from requirements.txt.
2. Fetch Dataset: Run python dataset_fetch.py to obtain and prepare the data. Data will take about 6GB of storage. Also, you may need to authenticate with your Kaggle account first.
3. Run Models:
   • For the CNN+LSTM approach: python cnn_lstm_approach.py
   • For the SVR approach: python svr_approach.py