Speech-based system for detecting alcohol intoxication using optimized deep learning

In this research, we propose a highly accurate speech-based system for detecting alcohol intoxication, utilizing deep learning algorithms to classify speech as either sober or intoxicated. The core of this system is the Dense Residual Recurrent Network (D2R_Net), which combines dense residual blocks for spatial feature extraction with Gated Recurrent Units (GRUs) for modeling temporal dependencies in speech data. This architecture is further optimized using the Iterative Parrot Optimization (ItPaO) algorithm, a novel hyperparameter tuning method designed to enhance the model's convergence and classification accuracy. The speech data is first pre-processed into log-Mel spectrograms, which allow the model to capture key auditory patterns altered by alcohol consumption. Through extensive experimentation, the proposed system outperformed existing models, with a balanced accuracy of 98.51% and specificity reaching 98.2%. The optimized model is highly efficient, requiring minimal computational resources while providing robust and scalable performance, making it ideal for deployment in real-world scenarios. This non-invasive, cost-effective alcohol detection system is poised to revolutionize public safety applications, such as monitoring drivers, event security, and healthcare settings.