Age Detection from Speech Signals

Politecnico di Torino
Data Science Lab Winter Project 2024/2025

---

1. Project Overview

This project focuses on building a regression pipeline to predict the age of a speaker based on their vocal characteristics. The task involves analyzing spoken sentences to extract acoustic and linguistic features that correlate with the speaker's age. The target output is a single numerical value representing the estimated age of the speaker.

---

2. Dataset Description

The dataset consists of 3,624 samples, divided into:

• Development set: 2,933 samples (with target age labels).
• Evaluation set: 691 samples (without target age labels).

Each sample corresponds to a spoken sentence, and the dataset includes various acoustic and linguistic features, such as:

• Acoustic features: pitch, jitter, shimmer, energy, zero-crossing rate, spectral centroid, tempo, harmonic-to-noise ratio, etc.
• Linguistic features: number of words, number of characters, number of pauses, silence duration.
• Metadata: gender, ethnicity, and file path to the audio recording.

---

3. Project Workflow

3.1 Data Exploration

• Inspected the dataset for missing values and outliers.
• Removed outliers based on the 99th percentile of the age distribution.
• Analyzed feature correlations and removed highly correlated features (e.g., num_characters and silence_duration).

3.2 Feature Engineering

• Extracted additional features from audio recordings, such as:
  • MFCCs (Mel-Frequency Cepstral Coefficients): Reduced dimensionality using PCA.
  • Spectral contrast and roll-off.
• Applied Within-Class Covariance Normalization (WCCN) to reduce intra-class variability.
• Standardized numerical features using z-score normalization.
• Encoded categorical features (e.g., gender and grouped ethnicities) using one-hot encoding.

3.3 Feature Selection

• Used a Ridge regression model to identify and drop features with low importance.

---

4. Model Development

4.1 Baseline Models

• Naive solution: Predicted the mean age for all samples.
• Simple linear regression: Used basic features for prediction.

4.2 Advanced Models

• Tested the following regression models:
  • Decision Tree Regressor
  • Random Forest Regressor
  • Support Vector Regressor (SVR)
  • Multi-Layer Perceptron (MLP) Regressor

4.3 Hyperparameter Tuning

• Performed grid search to optimize hyperparameters for:
  • SVR: Tuned C, epsilon, and kernel.
  • Random Forest: Tuned n_estimators, max_depth, max_features, and criterion.
  • MLP: Tuned hidden_layer_sizes, alpha, and learning_rate.

---

5. Evaluation

Metrics

• R² Score: Measures the proportion of variance explained by the model.
• Root Mean Squared Error (RMSE): Measures the average prediction error.

Results

• The best-performing models were:
  • SVR: Achieved the lowest RMSE with optimized hyperparameters.
  • Random Forest: Performed well with a balanced trade-off between accuracy and interpretability.
  • MLP: Demonstrated strong performance with deep learning techniques.

---

6. Predictions

Final predictions for the evaluation set were saved in the following files:

• TESTS/predictions_final_SVR.csv
• TESTS/predictions_final_RF.csv
• TESTS/predictions_final_MLP.csv

---

7. How to Run the Project

1. Install dependencies: Ensure you have Python installed along with the required libraries:

   pip install numpy pandas matplotlib scikit-learn librosa

2. Prepare the dataset: Place the dataset files (development.csv, evaluation.csv, and audio files) in the DSL_Winter_Project_2025/ directory.

3. Run the notebook: Open and execute the DSL-AgeDetection.ipynb notebook.

4. Generate predictions: The notebook will output predictions for the evaluation set in the TESTS/ directory.

Contributors:

• Maretto Chiara: s339401
• Gosmar Dario: s337625