DiaVoc

A machine learning system for predicting diabetes risk from voice recordings using federated learning and advanced audio feature extraction.

Overview

DiaVoc is an innovative AI-powered system that analyzes voice patterns to predict diabetes risk. The system leverages self-supervised learning with BYOL-S (Bootstrap Your Own Latent) for robust audio embeddings, combined with federated learning to ensure privacy-preserving model training across distributed clients.

Features

• Voice-Based Diabetes Prediction: Analyzes voice recordings to assess diabetes risk
• Federated Learning: Privacy-preserving training across multiple clients without sharing raw data
• Advanced Audio Processing: Uses BYOL-S embeddings for state-of-the-art voice feature extraction
• Web API: FastAPI-based REST API for easy integration
• Comprehensive Evaluation: Multiple ML models with thorough performance metrics
• Non-IID Data Handling: Robust partitioning for realistic federated scenarios

Project Structure

Voice-to-Diabetes-main/
├── src/                          # Core source code
│   ├── training.py              # Federated learning training pipeline
│   ├── infer.py                 # Inference system for predictions
│   ├── processing_pipeline.py   # Data preprocessing and feature engineering
│   └── generate_embedding.py    # Audio embedding generation
├── app/                         # Web application
│   ├── main.py                  # FastAPI server
│   └── inference.py             # API inference logic
├── fl/                          # Federated learning components
│   ├── fl_simulation.py         # FL simulation and aggregation
│   ├── data_loader.py           # Data loading utilities
│   ├── predict.py               # Prediction utilities
│   └── benchmark_models.py      # Model benchmarking
├── models/                      # Pre-trained models and artifacts
│   ├── serab-byols/            # BYOL-S audio encoder
│   ├── scaler.pkl              # Feature scaler
│   ├── pca.pkl                 # PCA transformer
│   └── global_model_improved.pkl # Trained global model
├── data/                        # Data files
│   ├── male_embeddings.pkl      # Male voice embeddings
│   ├── female_embeddings.pkl    # Female voice embeddings
│   └── *.wav                    # Sample audio files
├── README.md                    # This file
└── training_history.png         # Training visualization

Installation

Prerequisites

• Python 3.8+
• PyTorch
• CUDA (optional, for GPU acceleration)

Dependencies

Install the required packages:

pip install numpy pandas scikit-learn torch librosa fastapi uvicorn joblib matplotlib seaborn

Setup

1. Clone the repository:

git clone <repository-url>
cd Voice-to-Diabetes-main

2. Install dependencies:

pip install -r requirements.txt  # If available, or install manually as above

3. Ensure data files are in place:
   • Place male_embeddings.pkl and female_embeddings.pkl in the data/ directory
   • Ensure BYOL-S checkpoints are available in models/serab-byols/checkpoints/

Usage

Training the Model

Run the federated learning training pipeline:

python src/training.py

This will:

• Load and preprocess voice embeddings
• Simulate federated learning across 5 clients
• Train and aggregate models using FedAvg
• Save the global model to models/

Running Inference

For command-line inference:

python src/infer.py

Web API

Start the FastAPI server:

python app/main.py

The API will be available at http://localhost:8000

API Endpoints

• POST /predict: Predict diabetes risk from voice recording
  • Parameters: audio (file), age (int), gender (str), bmi (float), ethnicity (str)

Example usage:

import requests

files = {'audio': open('sample.wav', 'rb')}
data = {'age': 45, 'gender': 'male', 'bmi': 28.5, 'ethnicity': 'asian'}
response = requests.post('http://localhost:8000/predict', files=files, data=data)
print(response.json())

Data Preprocessing

To preprocess new data:

python src/processing_pipeline.py

This generates the necessary preprocessing artifacts (scaler, PCA) saved in models/.

Federated Learning Simulation

Run FL simulations:

python fl/fl_simulation.py

Methodology

Audio Feature Extraction

• Uses BYOL-S (Bootstrap Your Own Latent) for self-supervised audio representation learning
• Extracts 2048-dimensional embeddings from voice recordings
• Combines with demographic features (age, BMI, gender, ethnicity)

Federated Learning

• Implements FedAvg algorithm with true parameter averaging
• Non-IID data partitioning using Dirichlet distribution (α=0.5)
• 5 clients simulation with local training rounds
• Global model aggregation for improved generalization

Model Architecture

• MLP Classifier as primary model
• Feature engineering with Risk Index (Age × BMI interaction)
• PCA dimensionality reduction to 100 components
• Standard scaling for numerical features

Performance

The system achieves:

• 80%+ accuracy on held-out test sets
• Robust performance across demographic groups
• Privacy preservation through federated learning
• Real-time inference capability

Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests if applicable
5. Submit a pull request

License

This project is licensed under the MIT License - see the LICENSE file for details.

Citation

If you use this work in your research, please cite:

@misc{diavoc2026,
  title={DiaVoc: Voice-based Diabetes Prediction using Federated Learning},
  author={Khati, Harina and Pathak, Samriddha and Pokhrel, Jyoti and Subedi, Rasum},
  year={2026},
  howpublished={GitHub repository}}
}

Contact

Samriddha Pathak: +977 9702187444

Email: samriddhapathak123333@gmail.com