Development of an ai-driven secure safe box biometric authentication system using speech and face recognition technique

Abstract

The rapid advancement of Artificial Intelligence (AI) has enabled multimodal biometric systems that integrate speech and face recognition for more secure and reliable authentication. Unlike Unimodal systems that depend on a single trait and are prone to noise, lighting variations, spoofing, and user variability, multimodal systems combine complementary modalities to enhance robustness. In this study, Convolutional Neural Networks (CNN) were applied for face recognition and Recurrent Neural Networks (RNN) for speech recognition, with feature extraction using Mel-Frequency Cepstral Coefficients (MFCC), Linear Predictive Coding (LPC), and Perceptual Linear Prediction (PLP). A custom dataset was collected using ESP32-CAM for facial images and a CV-02 speech module for voice samples under varying environmental conditions. The CNN model achieved 93.5% accuracy, while the RNN gave 92.7% accuracy. When integrated to the app, the multimodal system significantly outperformed unimodal approaches with reduced False Acceptance Rate of 1.5% and the False Rejection Rate of 3.5%. These show that combining CNN and RNN models with advanced speech features provides robust, accurate, and secure real-time authentication, resilient to environmental and user-based variability.