Abstract
Person identification using breath sounds has emerged as a promising biometric modality. While previous research demonstrated the feasibility of using stethoscopes or single mobile devices, the critical challenge of cross-device generalization remains unexplored. This work presents the investigation of cross-device person identification using mobile-sensed bronchial breath sounds. We collected a multi-device dataset comprising recordings from 20 subjects using three different mobile devices, with three sessions per subject per device. We propose a framework combining the Audio Spectrogram Transformer (AST) with supervised contrastive learning and dual data augmentation (traditional techniques and diffusion-based augmentation). Evaluation using device-based and session-based cross-validation reveals that cross-device scenarios are significantly more challenging than same-device scenarios (a 5-7% gap in baseline). Our proposed method achieves 85.99-88.99% cross-device accuracy and 90.85-92.58% same-device accuracy, significantly narrowing the performance gap. The AST demonstrates clear superiority over CNN baselines (VGGish, ResNet50) with 16-25% advantages. Ablation studies confirm that supervised contrastive learning and data augmentation each contribute 2-4% improvement, with synergistic benefits when combined. Robustness evaluation shows the proposed method maintains 82-87% accuracy even at 0dB SNR, while the baseline degrades dramatically. Few-shot learning experiments demonstrate remarkable data efficiency, achieving comparable performance with only 5 segments per session. These findings establish the feasibility of practical cross-device breath-sound biometrics with well-acceptable noise robustness and minimal enrollment requirements.
Funding Information
  • National Science and Technology Council (NSTC 114-2221-E-027-039)

This publication has 17 references indexed in Scilit: