Infant Cry Verification with Multi-View Self-Attention Vision Transformers

This research addresses the novel challenge of verifying whether infant cries recorded at birth and discharge belong to the same infant. By utilizing Vision Transformers on MelSpectrograms with a multi-head self-attention mechanism, our method enables efficient feature extraction and verification. In a unique dataset preprocessing step, audio files from birth and discharge are merged into Mel-Spectrograms, where matched spectrograms are classified as belonging to the same infant (Identity 1), and non-matched spectrograms are classified as different infants (Identity 0). Inspired by models such as the Audio Spectrogram Transformer and ECAPATDNN, our approach is crucial for preventing infant mix-ups, ensuring data integrity, and tracking developmental health. This method demonstrates state-of-the-art performance on the Infant Cry Baby dataset, advancing AI-driven neonatal care.