Beyond fingerprint and voice print: cough sequence sound for identity verification

Traditional speaker verification systems rely on speech-based features such as fundamental frequency, resonance, and spectral characteristics. In clinical and remote monitoring settings where patients may be unable or unwilling to speak, non-speech sounds like coughs could provide a passive biometric signal for identity verification. While prior research has shown that individual coughs carry identity-specific features, most studies rely on single, voluntary coughs collected in controlled settings, limiting generalizability. Here, we investigated whether cough sequences contain sufficient information for speaker verification. We used the Coswara dataset, which contains 2746 cough recordings. The ECAPA-TDNN (Emphasized Channel Attention, Propagation, and Aggregation in Time Delay Neural Network) model, pre-trained on VoxCeleb, was adapted for cough-specific features. Speaker embeddings were extracted using SpeechBrain. Rejection rates were calculated by comparing shallow and heavy coughs from the same user, while false acceptance rates were computed by comparing a user's shallow cough with another's heavy cough. Equal Error Rate (EER) balanced false acceptance and rejection rates, while the F1 score measured precision and recall. We achieved an EER of 13.39% with an F1 score of 0.8559, indicating strong performance. Cosine distance between model embeddings effectively distinguished between genuine and impostor cough samples, confirming the model's ability to verify identity using cough sounds. We demonstrate the feasibility of using cough sequences for speaker verification, highlighting coughs as a viable signal for identity verification.