DSL-Age-Estimation

A Machine Learning project for the Data Science Lab: process and methods class, Age Estimation using both tabular and raw audio features. Developed in collaboration with my colleague Daniele Famà, this repository demonstrates the end-to-end process of data loading, feature extraction, preprocessing, and model training with advanced AI regression techniques.

---

Highlights

• AI Techniques: Utilizes CatBoost, Random Forest, and other ML algorithms for robust regression.
• Audio Feature Extraction: Gathers spectral, Mel-spectrogram, and time-domain statistics (skew, kurtosis, etc.) to capture essential speech cues.
• Tabular Feature Engineering: Cleans and preprocesses demographic and acoustic-linguistic data (outlier detection, scaling, encoding).
• Model Selection & Tuning: Employs dimensionality reduction (e.g., LDA) and hyperparameter search for optimal performance.
• Result: Achieved competitive RMSE scores on the public leaderboard, showing effectiveness of the combined approach.

For a detailed description of the methods, refer to the accompanying report.

---

Installation

1. Clone this repository:

   git clone https://github.com/AndreaLolli2912/DSL-Age-Estimation.git
   cd DSL-Age-Estimation

2. Install dependencies:

   pip install -r requirements.txt

---

Folder Structure

DSL-Age-Estimation
├── catboost_info
├── data
│   ├── audios_development
│   ├── audio_evaluation
│   ├── development.csv
│   └── evaluation.csv
├── figs
├── submissions
├── utils
│   └── audio_extraction.py
├── audio_features_preprocessing.ipynb
├── main.ipynb
├── report_figures.ipynb
├── Fama_Lolli_Age_Regression.pdf
├── README.md
└── requirements.txt

• data: Place all provided .csv files and audio folders here.
• utils/audio_extraction.py: Extracts raw audio features (adjust lowcut, highcut, top_db as needed).
• audio_features_preprocessing.ipynb: Converts raw audio features into tabular format (statistical summaries).
• main.ipynb: Runs the entire pipeline (EDA, model training, evaluation).

---

Usage

1. Extract Raw Audio Features

   python main/utils/audio_extraction.py

   Adjust cleaning parameters (e.g., lowcut, highcut, top_db) as desired.

2. Preprocess Features
   Open and run audio_features_preprocessing.ipynb to produce summarized data in .h5 or .csv formats.

3. Run the Model Pipeline
   Use main.ipynb to orchestrate data loading, EDA, and model training with CatBoost/Random Forest.

---

Credits

• Developed by @AndreaLolli2912 in collaboration with @danielefam.
• For full technical details, please refer to our report.

Feel free to open issues or submit pull requests for improvements.