Acoustic cues for person identification using cough sounds

• Enhanced approach to person identification using cough sounds (CPID) by improving model efficiency, robustness, and accuracy • Development of a compact deep learning architecture (CoughCueNet), trained with a hybrid supervised contrastive and cross-entropy loss to improve feature discrimination and generalization. • Thorough experimental evaluation confirms the effectiveness of the proposed system, achieving a high mean identification accuracy of 97.18% using MFCC features and data augmentation techniques. • Supervised contrastive learning significantly boosts model performance, outperforming conventional training strategies and larger models such as VGG16 and ResNet50 trained with cross-entropy alone • Strong potential for real-world applications, including user authentication, security, fraud detection, telemedicine, and public health monitoring—particularly in environments where speech-based systems are impractical. This study presents an improved approach to person identification (PID) using nonverbal vocalizations, focusing specifically on cough sounds as a biometric modality. While recent works have demonstrated the feasibility of cough-based PID (CPID), most report accuracies around 80-90% and could face limitations in terms of model efficiency, generalization, or robustness. Our objective is to advance CPID performance through compact model design and more effective training strategies. We collected a custom dataset from 19 subjects and developed a lightweight yet effective deep learning framework for CPID. The proposed architecture, CoughCueNet, is a convolutional recurrent neural network designed to capture both spatial and temporal patterns in cough sounds. The training process incorporates a hybrid loss function that combines supervised contrastive (SC) learning and cross-entropy (CE) loss to enhance feature discrimination. We systematically evaluated multiple acoustic representations, including MFCCs and spectrograms, to identify the most suitable features. We also applied data augmentation for robustness and investigated cross-modal transferability by testing speech-trained models on cough data. Our CPID system achieved a mean identification accuracy of 97.18%. Training the proposed CoughCueNet using a hybrid SC+CE loss function consistently improved model generalization and robustness. It outperformed the same network and larger-capacity networks (i.e., VGG16 and ResNet50) trained with CE loss alone, which achieved accuracies around 90%. Among the tested features, MFCCs yielded superior identification performance over spectrograms. Experiments with speech-trained models tested on coughs revealed limited cross-vocal transferability, emphasizing the need for cough-specific models. This work advances the state of cough-based PID by demonstrating that high-accuracy identification is achievable using compact models and hybrid training strategies. It establishes cough sounds as a practical and distinctive biometric modality, with promising applications in security, user authentication, and health monitoring, particularly in environments where speech-based systems are less reliable or infeasible.